Impact Evaluation of Tenondera - Paraguay’s Graduation Program

The intervention provides 2,864 households with a graduation program: a seed capital transfer to be used for the purchase of productive assets and a business and life skills coaching training curriculum.

2022-01-15

2024-12-31

Income, assets, savings, and mental health.

We will also explore heterogeneity of impacts treatment on these primary variables based on saturation rates in the neighborhood, baseline income and mental health and gender of household head.

Mental health will be measured using internal and external locus of control, self-efficacy, aspirations, and depression (adapted from Levenson's locus of control scale and the CES-D-10).

Time use, food security, mental health, women's empowerment, social capital.

Time use: number of hours of a typical day spent on work, household tasks, and caring for children or other people.

Food security: number of days in which adults over 18 skipped a meal in the past week, number of days in which children under 18 skipped a meal in the past week, number of days in which adults over 18 did not eat enough meat/fruits and vegetables/cereals and tubers in the past week, number of days in which children under 18 did not eat enough meat/fruits and vegetables/cereals and tubers in the past week (adapted from the Latin American and Caribbean Food Security Scale - ELCSA).

Women's empowerment: percentage of respondents with decision-making control over certain topics, percentage of respondents who agree that a male partner is justified in acting violently towards his female partner for certain reasons, level of agreement of respondents with traditional gender roles (adapted from Pro-WEAI, UNIANDES).

Social capital: percentage of respondents who are members of a group, percentage of respondents who describe their participation in a group positively, percentage of respondents who would speak in public in certain scenarios (adapted from Pro-WEAI, UNIANDES).

Randomization was conducted in two stages: at the neighborhood and at the household level. The randomization design ensured a close to even split of the households in the sample among three groups: early treatment, late treatment, and control. All control households will be offered the treatment after the end of the study.

First, the neighborhoods in the sample were split into a saturation design group and a non-saturation design group based on neighborhood population estimates. Those with a current population from 50 to 200 households were assigned to the saturation design group. Those with a current population below 50 households or above 200 households were assigned to the non-saturation design group. As explained below, it was felt that the saturation design would not work well in either low population or high population neighborhoods.

The non-saturation design group of neighborhoods was randomly split into two schemes: non-saturation scheme 1 and non-saturation scheme 2. In a second stage of randomization, 67% of households in each non-saturation scheme 1 neighborhood were assigned to early treatment and 33% to late treatment. 33% of households in non-saturation scheme 2 neighborhoods were assigned to late treatment and 67% to control.

The saturation design group of neighborhoods was split into six schemes in the first stage of randomization. In saturation scheme 1 neighborhoods, all households were assigned to early treatment. In saturation scheme 6 neighborhoods, all households were assigned to control. In the second stage of randomization, 80% of households in each saturation scheme 2 neighborhood were assigned to early treatment and 20% to late treatment, 20% of households in each saturation scheme 3 neighborhood were assigned to early treatment and 80% to late treatment, 80% of households in each saturation scheme 4 neighborhood were assigned to late treatment and 20% of control, and 20% of households in each saturation scheme 5 neighborhood were assigned to late treatment and 80% to control.

This decision to split neighborhoods into a saturation design group and a non-saturation design group was made based on estimated neighborhood eligibility rates, which we defined as the number of households in the sample from a given neighborhood over the estimated number of households living in that neighborhood. For us to gain insights into how our saturation design interacts with the community effects of the program, we need there to be a correlation between the variation in our imposed saturation rates in the sample and the saturation rates of the program in the population. Our concern was that high variation in eligibility rates across neighborhoods might have led to a breakdown of this correlation. That is, two neighborhoods with the same sample saturation rate but widely different eligibility rates will have two widely different population saturation rates, meaning our sample saturation rates would not be a good proxy of population saturation rates.

Therefore, we excluded neighborhoods with a small population from the saturation design since eligibility rates in those neighborhoods were much higher than average. Similarly, we excluded neighborhoods with a large population from the saturation design since eligibility rates in those neighborhoods were much lower than average. We did select neighborhoods with their current population within an interval around the mean to take part in the saturation design group since it is in those neighborhoods where the eligibility rates were restricted to a certain range.

When determining the different neighborhood schemes, our criteria were: ensuring a close to even split of the households in the total sample among the three treatment groups, having variation in the saturation rates, and having no neighborhood with some households assigned to the early treatment group and some households assigned to the control group. This third criterion was established to avoid potential political issues that may have arisen due to the large time difference at which different households in the same community would have been offered the program had we not prevented this kind of assignment.

Not available

Both neighborhood level and household level randomization were conducted in office using Stata.

First stage: neighborhood level. Second stage: household level.

No

246 neighborhoods.

2,864 households.

Neighborhoods: 59 in non-saturation scheme 1, 60 in non-saturation scheme 2, 22 in saturation scheme 1, 22 in saturation scheme 2, 21 in saturation scheme 3, 20 in saturation scheme 4, 21 in saturation scheme 5, 21 in saturation scheme 6.

Households: 946 in the early treatment group, 961 in the late treatment group, 957 in the control group.

At midline, we will be able to detect a mean change in income of 8.3% or higher. At endline, we will be able to detect a mean change in income of 9.5% or higher. In every case, the size of the minimum detectable effect is smaller than the one graduation programs commonly yield.