Information Frictions, Algorithmic Matching, and Worker Preference: A Two-Stage Randomized Control Trial on Turnover and Productivity in Manufacturing

We collaborate with the in-house factory of a Chinese technology firm manufacturing earphones. The factory currently faces a critical challenge: high turnover rates among general workers and a disconnect between worker capabilities and task assignment. The production line operates under a hierarchy (General Workers, Multi-skilled, Leaders), but recruitment is rapid and non-selective. Managers currently assign tasks based on limited heuristics (e.g., arrival order, gender, perceived energy), leading to potential mismatches.

To address these frictions, we propose a two-stage field experiment involving a projected cohort of 2,600 new hires. The design disentangles supply-side mechanisms (worker expectations and incentives) from demand-side mechanisms (managerial matching quality).

2026-03-07

2026-05-01

1. Retention: Survival analysis of tenure (days worked).
2. Productivity: Output per hour and quality control (error rates) normalized by task type.
3. Attendance: Absenteeism and lateness.
4. Compliance: The rate at which managers follow the AI recommendation or the Worker Preference (checking for principal-agent friction between researchers and floor managers).
5. Subjective Well-being: Self-reported job satisfaction and alignment with expectations (collected via follow-up surveys).

1. Worker Retention (Tenure)
Description: The duration a worker remains employed at the factory. This is the primary measure of whether the Hiring Stage treatments (expectations/incentives) and Job Assignment treatments (AI/Preference matching) successfully reduced turnover.
Construction:
Binary Retention Indicators: A set of binary variables equal to 1 if the worker is still employed at fixed intervals (e.g., Day 7, Day 14, Day 30, Day 90) and 0 otherwise.
Tenure Duration: The total count of days from the hire date to the separation date. For workers still employed at the end of the study window, this variable will be right-censored and analyzed using survival analysis models (e.g., Cox Proportional Hazards).
2. Standardized Productivity Index (Task-Adjusted)
Description: Since workers are assigned to different tasks (some harder, some easier) and different production lines, raw output counts are not directly comparable. We construct a standardized index to compare performance across the whole sample.
3. Assignment Compliance (Managerial Adherence)
Description: This outcome measures the extent to which managers utilized the information provided in the Job Assignment treatments (AI recommendations or Worker Preferences) versus their own discretion. This helps identify the "value of information" and principal-agent friction.
Construction:
Algorithm Compliance Rate: A binary variable equal to 1 if the worker's actual assigned task matches the task recommended by the AI algorithm (highest predicted score), and 0 otherwise.
Preference Compliance Rate: A binary variable equal to 1 if the worker's actual assigned task matches one of their "top 3" stated preferences collected during the hiring stage, and 0 otherwise.
4. Effective Labor Supply (Attendance & Overtime)
Description: Turnover is extensive margin (leaving), but productivity is also driven by the intensive margin (how much they work while employed).
Construction:
Calculated as the ratio of Actual Hours Worked (including voluntary overtime) to Standard Scheduled Hours.
This measure captures both absenteeism (ratio < 1) and willingness to work extra hours (ratio > 1), which is a proxy for worker engagement and motivation.
5. Expectation-Reality Gap (Belief Accuracy)
Description: Specifically for the Hiring Stage treatments, we measure if the video/incentive information reduced the "shock" of factory life.
Construction:
We calculate the absolute difference between Pre-entry Beliefs (surveyed at recruitment: expected hours, expected difficulty, expected earnings) and Post-entry Realization (surveyed at Week 1: actual experience).

This study employs a 3x3 cross-randomized design to disentangle the drivers of factory turnover and productivity through two distinct stages. In the Hiring Stage, recruits are randomized into a control group, a "Realistic Job Preview" group viewing a video on strict factory life to align expectations, or a "Task-specific Compensation Incentive" group receiving the video plus details on long-term wage incentives. Subsequently, in the Job Assignment Stage, managers assign these workers to tasks using one of three protocols: random assignment, an AI-assisted method utilizing algorithmic predictions of retention and productivity, or a Preference-based method incorporating workers' stated task desires. By interacting supply-side information treatments with demand-side matching mechanisms, the study tests whether retention is better improved by screening for fit and incentives or by optimizing the manager's assignment decisions through data and worker voice.

Not available

Randomization done in office by a computer

Individual

Yes

No

2000 newly hired workers

2600 newly hired workers

Minimum detectable effect size for main outcomes (accounting for sample design and clustering)