Food voucher and information for the TSA beneficiaries in Georgia

The participants in this study were households with children who receive the benefits from the Georgia Targeted Social Assistance Program (TSA). These households are selected using a household welfare index (proxy mean test - PMT) and receive benefits accordingly. Depending on the PMT score households receive a monthly transfer from GEL 60 to GEL 30 per individual. In addition, each child receive and extra transfer of GEL 50.

We first randomly divided these households in two groups. In the first group households receive their child benefit divided as GEL 20 in cash and GEL 30 as a food voucher. In the second group we delayed the implementation of the food voucher and give the child benefit GEL 50 in cash.

Currently the TSA program makes the total transfer per household via a deposit in a bank account. Those households in the voucher system receive an extra card where the GEL 30 per child could be used only in a list of grocery shops. In addition, the when the transfer is made at the end of each month, the household head receives and SMS informing that the TSA benefit is in his/her bank account.

We extended our study using an extra treatment. We added a third group which received GEL 50 per child in cash, but also received one messages (SMS) per month giving details about the transfer. Particularly, the message disentangles the transfer per household member and the transfer per child. We also added a forth group where we sent the SMS to households who were in the food voucher system.

To facilitate implementation, the random allocation was done at county (municipality) level. The final design is the following:

• Arm A (voucher): In these municipalities the implementation continues as planned and households with children will receive GEL 20 in cash and GEL 30 in the food card per child.
• Arm B (voucher + message): Individuals receive GEL 20 cash and GEL 30 voucher per child. Additionally, they receive one (or more) messages remembering that GEL 50 per child are benefits for children.
• Arm C (cash only): In these municipalities the new child benefit will be GEL 50 in cash per child.
• Arm D (cash + message): Individuals receive GEL 50 cash per child. Additionally, they receive one SMS per month remembering that GEL 50 per child are benefits for children.

* GEL 1 = USD 0.31

2019-03-01

2019-11-01

Our main outcome of interest is expenditure composition. Mainly, we aim to observe an increase in the share of food within the households budget. For the SMS recipients we aim to observe increases in the share of child related expenditure within the households (for example education and health).

We collected detailed expenditure information from a random sample of households.

We also collected information about children's health and education. Also information about parents' behaviour.
In addition we have access to administrative information about education (primary, secondary and tertiary), formal labor market participation and use of health services.

The allocation was done at municipality level. In order classify 64 municipalities in 4 groups and the level of comparability among arms we used the following process:
• We created 13 clusters using information about the number of beneficiaries (household with children and PMT lower 100,000), the total population and region.
• We aimed to have at least 4 municipalities per cluster. In three cases we were not able to create a cluster with 4 municipalities, for example Tbilisi was too different of all others and always created a cluster itself. Therefore, Tbilisi was not part of this study.
• Within each cluster, we randomly classified each municipality to one arm.

The resulting classification is described in the following table:

TABLE 1. ARMS COMPOSITION
Arm Municipalities Child benefit beneficiaries (households)
A 16 21,292
B 15 18,064
C 16 22,524
D 16 21,303
Source: Econometría S.A.’s calculations using PMT information. It does not include Tbilisi

As we can see, each arm has 16 municipalities and approximately 21 thousand households with child benefits. Additionally, for the impact evaluation household data collection we selected 36 municipalities belonging to all four arms.

Office by a computer

The unit of randomization was municipality (county).

Yes

64 clusters (municipalities).

We mixed administrative data with household data. These are the characteristics of the random sample of households for data collection. It is important to point out that the data collection was used for the evaluation of the TSA program by Regression Discontinuity Design. Therefore, the sample was taking around the cutoffs used by the program to target different benefits. ===== Definitions: PSU: Sampling areas: combination of city and zipcodes/village RDD: Regression discontinuity design PMT: Poverty score. Larger value means less poverty. Declaration: SA data collection in order to calculate the PMT TSA: Benefits Cutoffs: PMT values where the benefits change Experimental arms: We have 4 arms when we are texting voucher vs cash vs information. PPSWOR: Sampling without replacement with probability proportional to size. Sample selection: 1. Selection of the universe of analysis. a. Select only the households with children in the first interview. b. Drop all the PSU with less than 10 c. Drop all the individuals whose last status is inactive (source = 6) for reasons of eligibility (does not comply with all legal requirements) or change address. 2. Given that we are using a RDD. Given this we need to estimate the optimal areas of analysis. a. Estimate the expected benefit of a households depending on the household composition in the last interview and the PMT score in the last interview. b. Estimating the optimal bandwidth around each cutoff using the expected TSA benefit as outcome variable and the score at the first interview as score variable (rdrobust command in Stata) c. Given that some cutoffs are to close from each other we mixed the resulting optimal bandwidths with some restrictions. i. For the cutoffs at 30k and 100k we use the resulting optimal bandwidth. ii. We group individuals around 57k, 60k and 65k. We take individuals with scores from 54k to 70k (bandwidth of 3k). Then around 57k we have individuals from 54k to 60k. Around 60k we use individuals from 57k to 65k. Around 65k we use individuals from 60k to 70k (bandwidth of 5k). iii. We exclude all the households outside this areas. 3. Sample selection: a. Two conditions: i. The sample should allow to make estimations around each cutoff ii. Adding up the households by municipality, we should have enough households to make comparison between every two of our 4 arms (voucher/info/cash) iii. Tbilisi is not part of the experiment so we select households from Tbilisi only for the RDD analysis. b. Selection process: i. We first set an oversample of 20%. The final selection is 8494 households. ii. The selection was done in three stages, first municipalities, second PSUs, third households. iii. To select the municipalities we use PPSWOR sampling taking into account the number of households from each municipality in each experimental arm and around each PMT cutoff. iv. Within each selected municipality, PSU selection was done using PPSWOR sampling weighting each PSU according to the number of households in each PSU. v. In some cases we selected all the PSUs in the municipality. It depends on the size distribution within each municipality. vi. Within each PSU we selected households using Fan-Muller-Rezucha algorithm. vii. We use the same method for small and big PSUs. Selecting ad-hoc more households in big PSUs increases the variance of the estimations. viii. The resulting distribution of PSU size is described in the following table. The average PSU has 32 households, while the largest PSUs include more than 130 households.

ARM | Freq. Percent Cum.
A | 2,124 25.01 25.01
B | 1,673 19.70 44.70
C | 1,393 16.40 61.10
D | 1,848 21.76 82.86
Tbilisi | 1,456 17.14 100.00
---------+-----------------------------------
Total | 8,494 100.00

We expect a MDE of 8.1%