Hiding Identity

2025-06-24

2025-11-30

GUESS ACCURACY
Workers are asked to guess the caste category and jati of their partner, and to report how confident they are in the caste category guess. To avoid gaming, the guesses are not incentivised. Workers are also asked whether they think their partner will be able to guess their caste category correctly, allowing us to characterise whether workers have sophisticated beliefs about their ability to hide their caste.

PRODUCTIVITY / TASK (MIS)ALLOCATION
Number of fields entered (quantity), fraction correctly entered (quality), number correctly entered (mix of quantity and quality), identity of the partner that does the high status task (at task-by-day level).

SOCIAL RELATIONS
Incentivised WTP/WTA to work with the partner again vs. a new partner sharing the respondent's caste. Costly (in terms of effort) sign-up for a future shared meal with the partner. Workers are also asked whether they expect to meet up in future. Overlapping circles measure of identity fusion.

TRUST
Trust game played with partner, covering (i) amount to share if selected as sender, (ii) guess of how much partner shares if they are selected as sender, (iii) amount to return if selected as receiver (using the strategy method).

WELL-BEING AT WORK
Self-reports of how happy/self-conscious/confident/relaxed/anxious at work (answers from 0 = not at all to 3 = very). We have baseline measures of each to use as controls.

GUESS ACCURACY
When workers guess the caste of their partner, they are also asked to rank order the likelihood of each caste category.

[Amendment July 8, 2025, after N=70 endline surveys completed] For a richer measure of beliefs, we dropped this rank order measure in favor of an incentivised belief distribution measure, where the respondent allocates 10 tokens to caste categories based on how likely they think it is that their partner is in each category (following work by Adeline Delavande). To practice, participants first answer a similar question about the age of their partner.

GUESS MECHANISMS
Guesses of first and last name, reports of how they made their guess (e.g. by knowing the name, from physical appearance, etc.), reports of how they learned the name (if they say they know it), what they noticed about physical appearance (if it helped inform their guess).

PRODUCTIVITY / ALLOCATION
Self-report of happiness with how tasks were allocated (3-point scale: happy with allocation / would prefer to do more tablet work / would prefer to do less tablet work), report of whether they could have done better if participant had their way with task allocation. 10-point self-reports of own performance and partner’s performance.

SOCIAL RELATIONS
Each worker is asked to report how friendly / rude / bossy their partner was to them during the workday and how much tension they had with their partner, on a scale of 1 to 10. Self-report of whether able to show true self, and reports of what information about self was shared with the partner. For obfuscation of our caste-guessing questions: guess accuracy about age, marital status, and wealth of the partner, and how likely it is that the partner’s father works in a high-paid occupation.

MISCELLANEOUS
Open-text question asking what the participant thinks the study is about. Task-pair-level self-reported happiness and task difficulty.

We are recruiting high- (General caste) and low-caste (Scheduled caste) Hindu men with caste-distinctive last names in Uttar Pradesh. Participants complete a baseline survey when recruited at their home. Eligible participants are called in for one day of data entry work, with pay depending on performance and role assignment. There are three treatment groups, assigned with equal probability:

(i) HL-FullName: high-caste is paired with randomly-chosen low-caste; caste is made common knowledge through full names.
(ii) HL-FirstName: high-caste is paired with randomly-chosen low-caste; common knowledge of first names only.
(iii) HL-Hide: high-caste is paired with randomly-chosen low-caste; common knowledge of first names only; both workers are asked not to share their last name or other caste-identifying information with their partner.

We will use this design to answer two main sets of questions. First, we are interested in how successful attempts to hide caste are, and whether participants have sophisticated beliefs about whether others know their caste. For this we use endline questions that ask workers to guess the caste of their partner, and we see how guess accuracy is affected by the different treatment groups. In addition, we will explore (1) how guess accuracy compares with what an ML model can achieve using our baseline data, photographs, and voice recordings of each worker, (2) which baseline attributes (e.g. education, skin colour) predict more successful “passing,” and (3) which baseline attributes predict more successful guessing of the caste of partners.

Second, we are interested in how caste hiding affects task allocation, productivity, and social relations with the partner, with social relations captured by questions at endline. The data entry tasks themselves will vary in difficulty, with a subset of tasks carried out in random order. This allows us to test whether productivity effects depend on the nature and difficulty of the task.

For all tasks but one, pairs will also be asked to decide which worker does the “high-status” task of controlling the tablet, and which does the “low-status” task of reading the information to be recorded (from a set of printed sheets). The high-status task comes with a monetary bonus. We can then test for the effects of the treatments on task allocation. One task will be randomly chosen to have the high vs. low status role assignment randomly assigned.

For our analysis, we will run OLS regressions with randomization strata fixed effects, age and education, and a baseline-measured dependent variable when available. We will cluster standard errors at the pair-level (for outcomes measured at the individual level).

[Amendment January 13, 2025, when N = 58 had been completed] Our primary interest is in comparing endline outcomes between the three treatment groups. In the case that the caste-guessing accuracy is similar in the HL-Hide and HL-FirstName treatment groups, we will also pool HL-Hide and HL-FirstName for a higher-powered test of the effects of hiding caste.

[Amendment June 23, 2025] The intervention was initially launched in mid-January, as pre-registered; however, inspection of the outcome data revealed implausibly brief endline surveys recorded by the implementation partner, which made us concerned about data quality. As a result, we discarded the initial data and repeated the experiment from scratch. The new intervention start date is June 24, 2025. We updated this pre-registration on June 23rd. Other than changing the trial and intervention dates, the substantive changes made on June 23rd are:

(i) Modifying the reservation wage to get at working with one partner vs. another, rather than working vs. not working. With this change we aim to reduce noise in the revealed preference measure of enjoying working with the partner.
(ii) Adding a meal-sharing outcome, which was motivated by ongoing work on a separate project concerning untouchability practices.
(iii) Dropping the worker-pair photograph outcome given some resistance from pilot participants.

[Amendment July 8, 2025, when N=228 endline surveys completed] To improve the obfuscation of our design, we are now assigning one same-caste pair (HH or LL) to work together each day, always in the FullName condition. We alternate each day whether it is HH or LL.

Randomization will be done day-by-day using R code integrated with a shiny app interface. The randomization will be stratified by day-by-caste (where caste = high or low).

Individual-level

No

Approx. 1,000 to 2,000 individuals (500 to 1,000 pairs), depending on funding.

Approx. 1,000 to 2,000 individuals (500 to 1,000 pairs), depending on funding. Note that productivity data will be at the pair-level.

Each treatment: ⅓ of pairs