Care on Campus: Understanding and Responding to Student Wellbeing in India

This study evaluates MindMitra, an AI-powered mental health chatbot designed to reduce barriers to mental health care among undergraduate students in India’s public universities. The intervention is implemented as part of a randomized controlled trial across seven colleges within a large public university in Delhi. This is a web-based conversational chatbot that provides a private, low-cost, and low-friction entry point to mental health support. It is not a diagnostic or clinical tool. Instead, it is designed to help students reflect on emotions, normalize mental health concerns, reduce stigma, and provide information about professional mental health care.

We are running a follow-up intervention with 336 treatment group students who consented to MindMitra during the first round of deployment. We use these sign-ups to predict take-up in the control group. We keep the predicted users in the control group with above median prediction (logit). We then randomly select 336 control group students. We offer both the treatment and control group students a free introductory online session with three therapists.

The main intervention gives students access to an AI-powered chatbot called MindMitra (meaning “friend” in Hindi). This is a web-based conversational chatbot that provides a private, low-cost, and low-friction entry point to mental health support. It is not a diagnostic or clinical tool. Instead, it is designed to help students reflect on emotions, normalize mental health concerns, reduce stigma, and provide information about professional mental health care. Classrooms are randomized into either a control group or a treatment group. Students in control classrooms receive no access to the chatbot during the initial intervention period.

Students in treatment classrooms receive four weeks of access to MindMitra, beginning at the start of the spring semester. Each treated student receives a personalized login code. To encourage initial take-up and sustained engagement, all treated students are offered: 1) a sign-up incentive of INR 100 upon first registration, and 2) an engagement incentive of INR 100 for using the chatbot for at least 15 days.

Within treatment classrooms, students are further cross-randomized at the individual level into two additional intervention components designed to study mechanisms affecting adoption and engagement:
Alumni Endorsement (Role-Modeling Intervention)
Some students receive messages indicating that both male and female alumni from their university have used or support mental health care and therapy. These endorsements are intended to normalize help-seeking behavior, reduce perceived stigma, and lower the social cost of first use.
Engagement Streaks (Habit Formation Intervention)
Some students see a visible “streak” feature that tracks consecutive days of meaningful chatbot interaction (defined as more than one text per day). The streak is displayed in a scoreboard-style format to encourage repeated engagement and habit formation.
Students may receive neither, one, or both of these features depending on random assignment.

The intervention lasts for an initial four-week period, at the end of which students complete a midline survey. Following midline data collection, treatment-group students regain access to MindMitra for the remainder of the semester, allowing the study to observe longer-term engagement patterns and downstream outcomes.

2026-02-04

2026-03-11

Primary outcomes include adoption and engagement with MindMitra, measured by login, frequency of use, number of messages, and return rates. Additionally, we focus on mental health outcomes using psychometric tests (PHQ-4, GAD-4, UCLA-3) and mental health care-seeking behavior, including actual and intended use of professional mental health services (therapists or counselors).

For the follow-up therapist intervention, we are interested in the following outcomes:
1. Do they click on the Calendly links of any therapist (we will include this as an outcome variable if the majority of students' email accounts do not block this feature)?
2. Do they book a session?
3. Do they show up for the booked session?

During the endline survey, we plan to collect follow-up information post-intervention:
1. Care-seeking: used professional mental health services, use of crisis helplines, willingness to see a therapist in the future, and changes in self-reported barriers to mental health care.
2. Therapist intervention follow-up that measures self-reports of a) received voucher b) signed up c) counselor showed up d) repeat bookings, if any
3. Willingness to pay for formal therapy and future use of MindMitra.
4. We have added the WHO-5 Well-Being Index to capture any positive changes in mental health outcomes. This is because the existing battery of questions only captures negative experiences.

We will create indices for some of the mental health outcomes variables. This includes a Wellbeing Index, which combines the PHQ-4 and GAD-4 outcomes as is prevalent in the literature. We will also convert the other Likert-style questions into indices (GLS indices, Anderson (2008)) wherever feasible.

Secondary outcomes include own and perceived stigma toward peers seeking mental health care. We measure willingness to pay for therapy through a donation game and a discrete choice experiment. We will also measure trust in and acceptance of AI for mental health support. Lastly, we measure changes in gendered agency around marriage expectations, ability to push back on parental expectations, and future academic and job market aspirations.

During the endline survey, we plan to collect follow-up information post-intervention:
1. Own stigma, perceived judgment from others, and perceived peers' views of both therapy and MindMitra use (second-order beliefs).
2. Trust in and acceptance of AI for mental health support, which is similar to the questions asked at baseline and midline.
3. Referral behavior: whether participants invite classmates to try MindMitra, how many they refer, and reasons for declining to refer. We randomize participants into whether the referral is anonymous versus named to understand if stigma around named referrals reduces the likelihood of inviting more students to use MindMitra. This is purely for experimental purposes and we will not actually expand MindMitra access at this stage (and participants are debriefed on this during the survey).
4. Civic/advocacy behavior: willingness to sign a petition calling for expanded on-campus mental health resources, as a behavioral proxy for reduced stigma and demand for support. Participants can opt to sign this petition anonymously or with their names. This is hypothetical and participants are debriefed on this at the end.
5. Grit and self-regulation (persistence and consistency of effort toward long-term goals).
6. General self-efficacy (perceived ability to cope with and manage difficult situations).
7. Changes in gendered agency around marriage expectations, ability to push back on parental expectations, and future academic and job market aspirations.
8. Academic performance: self-reported GPA (verified via transcript upload) and perceived relative academic performance compared to peers.
9. Academic functioning: class attendance, attention/concentration difficulties, mind-wandering, and academic stress interfering with sleep or daily functioning.
10. Social media and screen time use, including attempts to cut back and time spent across platforms.
11. Mental health literacy and self-care practice behaviors (e.g., sleep, exercise, gratitude, social connection, screen time limits).
12. Demand for future on-campus wellness programming (interest in hypothetical free activities such as peer support, group fitness, or mindfulness sessions).

We will convert the other Likert-style questions into indices (GLS indices, Anderson (2008)) wherever feasible. This includes responses for grit, self-efficacy, stigma, AI acceptability, academic functioning, and marriage market agency.

This study is a randomized controlled trial among undergraduate students in public universities in Delhi. Classrooms are randomly assigned to either a control group with no access to the intervention or a treatment group that receives access to an AI-powered mental health chatbot (MindMitra) for four weeks. Within treated classrooms, students are cross-randomized at the individual level to receive alumni endorsements, engagement streak features, both, or neither. Outcomes are measured using baseline, midline, and endline surveys, along with chatbot usage data.

We move to an individual randomization for the follow-up therapist intervention.

We randomize using Stata.

Randomization is at the cluster level (classrooms) as well as individual level within the treatment clusters.

Yes

233 classrooms across 7 colleges and 9 majors.

4,452 students. We rely on self-administered surveys and will do our best to reach all students who filled out the Baseline survey. However, some students have shared fake email addresses and phone numbers, which has led to uncertainty about the final sample size. The final sample size after the midline survey is at 4071. This excludes surveys with no viable contact information (including false emails and missing phone numbers).

90 control classrooms and 143 treatment classrooms. For the follow-up therapist intervention, we use the number of sign-ups (336) in the main intervention and select an equal number of students from the control group.

The study is powered at 80% with a 5% significance level, accounting for classroom-level clustering. For primary care-seeking outcomes, the minimum detectable effect is approximately 4 percentage points, corresponding to about 0.08 standard deviations. For depression outcomes, the study can detect a change of approximately 0.5 points on the PHQ scale.