Government Scheme Overload: Take-Up and Targeting in a Multi-Scheme Context

I implement a randomized experiment that varies the quantity and quality of information shared to potential beneficiaries about up to 15 priority social protection schemes. In the status quo, frontline workers approach individuals in communities, ask a series of demographic questions to understand their scheme eligibility, then provide individuals with information on the programs they are eligible for. This process is currently ad hoc and varies by frontline worker. My intervention involves testing different information sharing protocols frontline workers follow in the information sharing stage.

I vary quantity through the sequencing and presentation of information. Specifically, I vary whether or not citizens learn about one scheme at a time with opt-in continuation versus all eligible schemes at once. I vary quality by sharing information about random or recommended schemes. Taken together, I define three intervention arms: a single-random scheme arm, a single-recommended scheme arm, and a multiple scheme arm.

2026-09-15

2026-10-30

I define three primary outcomes to measure take-up quantity and quality. I measure each outcome at the end of a fixed interaction window (1 week from first interaction).

(1) Outcome 1: Quantity: The total number of schemes the individual applied to (between 0 and 15 schemes, conditional on eligibility)
(2) Outcome 2: Continuous Quality: The total predicted quality of scheme applications, defined as the sum of the quality indexes for each scheme an individual applied to. Since each measure of scheme quality is standardized, applying to low-quality schemes reduces the outcome, while applying to high-quality schemes increases it. Individuals who apply to 0 schemes receive a value of 0 for this outcome.
(3) Outcome 3: Targeted Quality: A binary indicator for whether or not the individual applied to the scheme with the highest predicted quality. Individuals who apply to 0 schemes receive a value of 0 for this indicator.

Measuring Take-up Quantity (used in Outcome 1)

I measure the quantity of take-up using scheme applications. I prioritize measuring applications over availed schemes because it better models an individual's decision. The 15 priority schemes I selected have historically not had many supply-side challenges, but their processing timelines differ widely (some schemes might issue immediately, while others might take months to process). Therefore, I measure take-up using applications as it captures the demand-side decision individuals make, and can be consistently measured across schemes at a fixed period of time.

Measuring Take-up Quality (used in Outcomes 2 and 3)
I measure take-up quality by constructing an index that proxies the net value of a scheme. I construct this index prior to the start of the intervention, then use a machine learning model to predict scheme quality for each individual who applies to schemes during the intervention period.

- Pre-intervention index construction: I collect survey data from individuals that interacted with Haqdarshak in Haryana in 2025. The phone survey asked beneficiaries about the 15 priority schemes they were eligible for. For each scheme an individual self-reported having availed, individuals were asked about the scheme's benefits (helpfulness, benefits relative to expectations, counterfactual reliance), application costs (documentation burden, office visits, monetary costs), demand (usage, probability of use), and monetary evaluation (willingness to accept giving up a scheme). Citizens who had not availed certain schemes they were eligible for were asked parallel questions about reasons for non-application and their perceptions of each scheme's benefits, application costs, and demand. I combine this survey data with administrative application data and scheme-level attributes to construct a menu of potential quality indices.

I select the final index through a treatment-blind procedure prior to the start of the evaluation. I review each candidate index and run two main checks. First, I check each index's variation. The final index should show meaningful variation across schemes and individuals. If an index does not vary much across either dimension, it will not be able to be ranked. Second, I check each index's consistency with other benchmarks. In the phone survey, I ask individuals about their most helpful scheme and overall welfare. Indexes should be consistent with a held-out revealed preference benchmark. Overall, using pre-intervention survey variation, I choose an index that measures scheme-quality for each individual that has high variance and is consistent. Since I select this index prior to the intervention, it is not influenced by the RCT's results. I standardize the final quality index prior to the RCT.

- Pre-intervention predictive modeling: I train a supervised machine learning model to predict the quality index for each individual-scheme based on that individual's baseline demographic characteristics. I freeze this model prior to the RCT.

- Intervention: For each individual a frontline worker interacts with during the RCT, I predict scheme quality for each scheme an individual is eligible for using the machine learning model that was trained on pre-intervention data. Overall, my metric for scheme quality was designed using a data-driven approach pre-intervention.

I define a set of exploratory outcomes to understand the mechanisms behind the take-up decisions individuals make when exposed to different quantities and qualities of schemes. First, I explore the extensive margin and timing of applications: an indicator for any scheme application, each primary outcome measured ``on-the-spot'' during the first interaction with a frontline worker, each primary outcome measured during subsequent interactions, the number of subsequent interactions, and the total time in between initial and subsequent interaction. Second, I explore the demand for information on schemes by measuring the number of schemes an individual applied to before stopping, an indicator for stopping after the first scheme, and measures of revealed preferences. Third, I explore schemes availed by using the same primary outcomes with scheme receipt instead of simply applications. Conditional on additional funding, I can conduct a follow-up endline survey to measure self-reported welfare. Finally, I can define binary indicators for take-up at a scheme level to understand the within-scheme effect of each intervention.

- Arm 1: Single-random scheme - Frontline workers approach a citizen, introduce themselves, introduce Haqdarshak, and offer to help connect citizens with social protection schemes. Once citizens consent, frontline workers ask citizens a series of demographic questions and fill out a mobile application. The mobile application then identifies which schemes the citizen is eligible for. For each citizen interaction, frontline workers receive the list of schemes in a randomized order. Frontline workers share information about the first scheme on the list the citizen is eligible for: the scheme's benefits, the documentation required to apply, and the steps required. If the citizen is interested in the scheme, the frontline worker assists them in their application. The citizen can apply ``on-the-spot'' if they have the relevant documentation with them, or they can coordinate another time to meet with the frontline worker and apply. After applying for the first scheme, frontline workers ask the citizen if they would like to learn about another scheme. If citizens are interested, the frontline worker shares information about the next eligible scheme. In this way, frontline workers share information on eligible schemes sequentially instead of all at once.

- Arm 2: Single-recommended scheme - Frontline workers follow similar protocols to Arm 1, except instead of providing information on a random scheme, they provide information on a scheme they recommend. Frontline workers leverage their expertise and have autonomy to provide information on the single ``best'' scheme per interaction.

- Arm 3: Multiple schemes - Frontline workers approach and interact with citizens in a similar manner as in other arms. However, instead of sharing information on schemes sequentially, they share information on all 15 priority schemes the citizen is eligible for at once. This means citizens may receive information on up to 15 schemes. The order of the schemes shared is random for each citizen.

Across Arm 1 and Arm 2, the quantity of schemes shared is fixed (one scheme per interaction), but the quality of the scheme shared differs. Differences between these groups in terms of the quantity of schemes applied to highlight the effect of availing schemes with a high net value. Across Arm 1 and Arm 3, the quality of schemes shared is fixed via random ordering, but the quantity of schemes shared differs. Differences between these groups in terms of the quality of schemes applied to highlight how complexity (proxied by information quantity) impacts take-up.

Not available

I randomize on two levels: geographic units and frontline workers. My sample contains 30 villages in Haryana, with each village divided into 3 geographically distinct units. I construct a grid of six clusters and three time periods, creating 18 cluster-periods. These clusters represent the six order permutations in which the three treatment protocols can be implemented across three time periods. This ensures that arm assignment is orthogonal to period by construction. In a given period, each frontline worker works in one geographic unit. Therefore, crossover operates at a frontline worker level (each frontline worker implements all three arms across all periods), not an individual level. Individual citizens only interact with one frontline worker during the period the worker was assigned to their geographic area. This ensures that each citizen is only exposed to one intervention protocol, meaning there is no within-citizen carryover. Through this design, I partial out the quality of each frontline worker in scheme delivery, which is important given variability in performance.

Randomization is done via a computer.

The geographic unit level. There are three distinct geographic units per village. In total, I will have 90 geographic units. Each unit will eventually be assigned to a treatment arm, frontline worker, and time period based on the crossover randomized design. I measure outcomes at the individual level for individuals in each geographic unit who interacted with a frontline worker assigned to a particular treatment arm protocol.

Yes

90 geographic units
(Each geographic unit is at the sub-village level. There are 3 geographic units per village and 30 villages total).

2,000 individuals

30 geographic units assigned to treatment arm 1 (single-random-scheme)
30 geographic units assigned to treatment arm 2 (single-recommended-scheme)
30 geographic units assigned to treatment arm 3 (multiple-scheme)

I use monte carlo simulations to simulate power. At N = 2,000, a minimum detectable effect of 0.1 total scheme applications yields power = 0.89.