The Erosion of Agency: How Automated Financial Nudging Impacts Long-term Economic Resilience in Digital Finance.

How to cite

Lopez, Francisco Jacobo Murillo. 2026. "The Erosion of Agency: How Automated Financial Nudging Impacts Long-term Economic Resilience in Digital Finance.." AEA RCT Registry. AEARCTR-0018964. https://doi.org/10.1257/rct.18964-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Francisco Jacobo Murillo Lopez*

* Primary investigator

Registry ID
AEARCTR-0018964
DOI
https://doi.org/10.1257/rct.18964-1.0
First registered
Last updated
Status
In development
Trial dates
2026-07-01 – 2027-01-05
Countries
Mexico (Aguascalientes)
JEL codes
D14, D91, G51, G53, O33, C93
Registry record
http://www.socialscienceregistry.org/trials/18964

Abstract

This study investigates whether delegating financial decisions to automated algorithms (automated nudging) erodes human financial agency and compromises long-term economic resilience against unexpected macroeconomic shocks. Grounded in behavioral economics and economic psychology, the research explores the "Algorithmic Delegation Paradox," examining if automation shifts users from empowered decision-makers to passive observers, thereby weakening their organic capacity to adapt to financial stress.

We implement a prospective, longitudinal randomized controlled trial (RCT) with a three-arm individual-level randomization structure: (A) Manual Management (Control), (B) Hybrid/Nudge Management (Active choice architecture), and (C) Fully Automated Management (Algorithmic delegation). Over a six-month period, participants' real-time behaviors, cognitive loads, and financial self-efficacy will be tracked using the Experience Sampling Method (ESM). To evaluate long-term resilience under pressure, two unexpected exogenous price shocks (inflationary simulations) will be introduced mid-experiment.

Primary outcomes include the accuracy of financial balance awareness (perception error), response latency to financial stress, and budgetary adjustment capacity (measured via the Shadow Liquidity Index 2.0). Data will be analyzed using fixed-effects panel models, Latent Growth Curve Models (LGM) to chart temporal trajectories, and Causal Forests to uncover treatment heterogeneity among vulnerable subgroups. This trial aims to provide critical empirical evidence for regulators, Central Banks, and international organizations designing consumer protection frameworks for the digital finance era.

Intervention

We implement a prospective, individual-level randomized controlled trial (RCT) with a longitudinal design spanning six months (July–December 2026). The study evaluates the behavioral and cognitive impacts of algorithmic choice architecture in digital finance, specifically examining how varying degrees of automated financial decision-making (automated nudging) affect personal financial agency and long-term economic resilience.

A sample of N=750 participants will be selected and randomly assigned to one of three experimental arms (1:1:1 allocation ratio), stratified by baseline age and educational attainment:

Arm A: Manual Management (Control Group): Participants manage their personal budgets and liquidity entirely manually through a standardized digital finance interface. They receive standard transaction records and balance updates but do not receive automated behavioral nudges, suggestions, or autonomous execution features. This arm measures baseline financial agency and organic adaptation.

Arm B: Hybrid/Nudge Management (Active Choice Treatment): Participants interact with an interface embedded with real-time digital behavioral nudges (e.g., proactive alerts on spending thresholds, reminders of financial goals, and context-dependent choice framing). Crucially, the system requires active user intervention and explicit confirmation to execute any budget reallocations or micro-savings adjustments, thereby preserving cognitive engagement and the decision-making loop.

Arm C: Fully Automated Management (Algorithmic Delegation Treatment): Participants delegate operational financial control to an algorithmic optimizer. The system autonomously executes micro-savings, schedules transfers, and reallocates discretionary spending based on predefined parameters without requiring step-by-step user confirmation or active choice. This arm isolates the impact of "automated nudging" and structural algorithmic reliance.

Longitudinal Monitoring & Exogenous Shocks:
Over the 24-week period, high-frequency granular behavioral data and subjective financial states will be captured twice weekly using the Experience Sampling Method (ESM) via mobile devices. To empirically assess resilience, two unexpected, exogenous macroeconomic price shocks (staged inflationary simulations that increase utility costs or essential transaction fees within the platform's simulated ecosystem) will be introduced: the first at the end of July (Week 4) and the second in September (Week 12). These shocks will isolate differences in adaptation latencies, perception errors regarding account balances, and real-time budgetary adjustments across the three arms.

Experimental design

Experimental Design Structure

1. Type of Study and Allocation Framework:
This study is structured as a prospective, longitudinal, individual-level randomized controlled trial (RCT) spanning 24 weeks (6 months) from July to December 2026. Randomization will be executed with a 1:1:1 allocation ratio across three independent experimental arms: Arm A (Manual Management / Control), Arm B (Hybrid/Nudge Treatment), and Arm C (Fully Automated Management / Treatment).

2. Randomization Strategy:
Randomization will occur at the individual level immediately following the baseline data collection (T0). To ensure covariate balance across the experimental arms, we will implement a stratified block randomization strategy. The stratification factors are baseline age cohorts and educational attainment levels, which are critical predictors of financial literacy and digital interface adoption. The randomization sequence will be generated via a computer algorithm in RStudio, ensuring full concealment from both the participants and the field researchers until the moment of assignment.

3. Sample Size, Power Analysis, and Attrition Target:
Based on an a priori power calculation for longitudinal panel models, we established a target power of 80% (1 - Beta = 0.80), a significance level of 5% (Alpha = 0.05), and a Minimum Detectable Effect Size (MDES) of f^2 = 0.15. The statistical power framework requires a minimum of 160 active participants per experimental arm, equating to 480 effective subjects. To aggressively safeguard against potential panel attrition, the trial will recruit an initial sample size of N = 750 participants. This structure absorbs a projected 30% attrition rate over the 6-month timeline while retaining approximately 500 final effective participants, comfortably exceeding the power threshold. Furthermore, the high-frequency measurement design (48 total points per individual via twice-weekly Experience Sampling Method prompts) drastically enhances intra-individual precision and statistical power.

4. Econometric and Analytical Strategy:
The public data analysis plan relies on three highly parsimonious and complementary analytical frameworks to evaluate the temporal dynamics of the intervention:
* Fixed-Effects (FE) Panel Models: Used as the primary econometric strategy to control for time-invariant unobserved individual heterogeneity, isolating the clean causal impact of automated nudging over time.
* Latent Growth Curve Models (LGM): Deployed to parametrically model and map the individual and aggregate longitudinal trajectories of financial agency and resilience across the 24-week period, integrating the exogenous price shocks as time-dependent predictors.
* Causal Forest (Machine Learning Causal Inference): Utilized to optimize the estimation of Conditional Average Treatment Effects (CATE). This approach will systematically uncover treatment effect heterogeneity and identify specific vulnerable socio-economic or cognitive subgroups without risking arbitrary post-hoc specification mining.

Other complex structural or multi-group extensions, such as Structural Equation Modeling (SEM) for latent construct validation, will be restricted strictly to secondary verification to preserve model parsimony. Non-random attrition patterns will be verified using the Verbeek-Nijman test, complemented by Inverse Probability Weighting (IPW) adjustments if necessary. Potential extreme data volatility in behavioral latencies or expenditure entries will be managed using systematic winsorizing at the 1% and 99% thresholds.

Design details

Randomization method
The individual-level randomization will be conducted in-office by a computer using a reproducible pseudo-random number generator algorithm. Specifically, the block-stratified randomization sequence will be pre-programmed and generated in RStudio using a secure seed for reproducibility. The final allocation to the three experimental arms (Arm A, Arm B, or Arm C) will be dynamically executed via an automated script embedded within the digital finance platform's backend infrastructure. This process occurs automatically and immediately after a participant's baseline (T0) demographic and psychometric survey is completed and locked, ensuring complete concealment from both field researchers and participants.
Randomization unit
Individual. Randomization is executed strictly at the individual level. Each participant who completes the baseline assessment is independently and randomly assigned to one of the three experimental arms (Arm A, Arm B, or Arm C). There is no cluster-level or multi-level hierarchical randomization in this trial design.
Clusters
Not applicable. This study utilizes an individual-level randomization design and does not contain clusters.
Observations
750 individuals (initial planned sample size). Given the longitudinal design of this trial, which captures high-frequency panel data over a 24-week period using twice-weekly Experience Sampling Method (ESM) prompts, each participant is expected to contribute up to 48 distinct measurement points. Therefore, the maximum planned number of longitudinal observations under zero attrition is 36,000 observations (750 individuals * 48 measurement points). Accounting for our conservative 30% projected attrition rate, we anticipate a minimum effective final sample of approximately 500 individuals, yielding an effective dataset of at least 24,000 panel observations for the final fixed-effects and latent growth curve analysis.
Arms
250 individuals assigned to Arm A: Manual Management (Control Group). 250 individuals assigned to Arm B: Hybrid/Nudge Management (Active Choice Treatment Group). 250 individuals assigned to Arm C: Fully Automated Management (Algorithmic Delegation Treatment Group). Total planned sample size: 750 individuals. This individual-level allocation ensures a perfect 1:1:1 balanced distribution at baseline across all three experimental arms.
Primary outcomes
We define three primary, high-frequency outcome variables (endpoints) to measure the structural erosion of financial agency and economic resilience: 1. Balance Awareness Error (Perception Deviation): Measured as the absolute mathematical difference between the participant's subjectively reported account balance during an Experience Sampling Method (ESM) prompt and their true, objective system balance at that exact timestamp: Error = Absolute_Value(Reported_Balance - True_Balance). This continuous variable operationalizes financial salience and cognitive detachment induced by automated delegation. 2. Response Latency to Financial Stress Events: Defined as the exact time elapsed (measured in seconds via system logs) from the precise moment an automated platform alert or exogenous shock notification is pushed to the participant's mobile device, until the user actively engages with the interface to review, override, or acknowledge the financial event. This metric captures behavioral inertia, automation bias, and cognitive friction under pressure. 3. Real-time Budgetary Adjustment Capacity (Shadow Liquidity Index 2.0): A composite operational index that aggregates a participant's high-frequency spending contractions, reallocations toward non-discretionary commitments, and active micro-saving preservation behaviors immediately following the two introduced exogenous price shocks. It quantifies organic economic adaptation versus artificial stability.
Secondary outcomes
We define three continuous and categorical secondary outcome variables (endpoints) to capture cognitive, psychological, and behavioral mechanisms underlying the intervention: 1. Financial Self-Efficacy Score: A continuous measure adapted from Bandura's self-efficacy frameworks, tracking the participant's subjective belief in their capability to execute autonomous financial planning and withstand unexpected economic strain. 2. Subjective Cognitive Load Score: A continuous index adapted from the NASA-TLX scale, measuring the mental effort, frustration, and psychological friction experienced by the participant during financial interface interactions and shock events. 3. Price Recall Accuracy (Financial Memory): A continuous metric measuring the absolute mathematical percentage error between a participant's recalled transaction price or current essential goods cost and the objective historical system logs. This isolates the "Financial Cognitive Atrophy" effect across the arms.

Cited by

No citation data available for this registration yet.

Keywords