Women in Tech: A Cluster-Randomized Field Experiment in Armenian Secondary Schools

2026-09-14

2026-11-30

Tech choice; Tech/AI fields more appropriate for men; AI-Job Self-Efficacy

Tech choiceᵢ: indicator that a tech field was selected in Q1. For definition of tech field see section 5.1 of analysis plan.
Tech/AI fields more appropriate for men: The 1-10 response to question Q8 for tech fields (see definition in section 5.1) of analysis plan) and AI fields (Q8.j).
AI-Job Self-Efficacy: The 1–5 response to Q9f ("I would perform well working in AI-focused jobs").

• SO1 and SO2 —The 1-5 answers in Q7 for “jobs in the sector involve working for people” and “good and services created in the sector improve people wellbeing”, averaged within the same respondent, for tech fields (SO1) and AI (j) (SO2).
• SO3 and SO4 — Work-life-balance index, role-model-specific: the 1-5 answer in Q7 from the work-life-balance item for tech fields (SO3) and (j) (SO4).
• SO5 —The 1–5 response to Q9e ("I would perform well working in math-focused jobs").

The intervention is a four-module classroom-based career-guidance program ("Career Guidance Program Promoting Gender Balance, STEM-Oriented"), developed to be consistent with the Ministry of ESCS’s pilot standard for the "Career Guidance" course (grades 8–11) and delivered by professional career-orientation trainers. Total program duration is 180 minutes, split across four modules:
• Module 1 — Self-Awareness and Discovery of Interests (introductory). A neutral module in which students complete a validated vocational-interest inventory (Holland/RIASEC) and reflect on their own interests and strengths. Delivered to all three arms (Control, Treat 1, Treat 2).
• Module 2 — My Values and Stereotypes. Students work through exercises eliciting and critically examining gender stereotypes in career and profession choice.
• Module 3 — Meeting with a STEM Specialist (role-model module). A woman working in a technical/AI role at a Yerevan AI or technology company (e.g., Krisp, Podcastle, Diagnocat, SuperAnnotate, Cognaize, YerevaNN, NVIDIA Yerevan, PicsArt) visits the classroom to describe her career path and current work, using a structured role-model protocol. Role models are recruited through the research team’s professional networks, LinkedIn outreach, and referrals from AI companies operating in Yerevan.
• Module 4 — Action Planning (closing session, not evaluated). Students complete an individual career-planning exercise. Delivered to Treat 1 and Treat 2 only, after both content modules; not designed to be separately evaluated and excluded from the primary and secondary outcome analysis.
The two active treatment arms receive the identical set of modules but in reversed order, which is the key design feature enabling separate identification of the role-model module from the values-and-stereotypes module:
• Treat 1: Baseline survey → Module 1 → Module 2 (Values and Stereotypes) → [Survey 2] → Module 3 (Role model) → Module 4
• Treat 2: Baseline survey → Module 1 → Module 3 (Role model) → [Survey 2] → Module 2 (Values and Stereotypes) → Module 4
• Control: Baseline survey → Module 1 only → [Survey, timed to match Survey 2]
Module 1 is delivered approximately 5 days after the baseline survey; the first content module (Module 2 for Treat 1, Module 3 for Treat 2) follows approximately 3–5 days later. Survey 2 (the main evaluation, see Section 3.4) is administered immediately after this first content module and before the crossover module is delivered, so that at the time of Survey 2, Treat 1 has received Module 1 + Module 2 only and Treat 2 has received Module 1 + Module 3 only — allowing the T1-vs-Control and T2-vs-Control contrasts to isolate the effect of Module 2 and Module 3, respectively. The crossover module is delivered immediately after Survey 2, and the closing Module 4 follows approximately 8 days later, both for Treat 1 and Treat 2.
The intervention is evaluated in a difference-in-differences design, comparing students in treated and control classes at endpoint vs baseline survey.

Not available

Treatment will be randomized at the class level, separately within each school. Each participating school includes three classes, with exactly one class assigned to the control group, one to Treatment 1 (T1), and one to Treatment 2 (T2). Randomization will be implemented by generating a uniformly distributed random number for each class, sorting the three classes within each school according to this number, and assigning the class with the lowest random number to Control, the second to T1, and the third to T2. A fixed random-number seed will be set before randomization to ensure full reproducibility. Prior to generating random numbers, the dataset will be sorted by school and class identifiers to ensure that the randomization is exactly reproducible from the recorded seed and dataset.

The procedure can be implemented in Stata as follows:

* Put observations in a unique order
sort school_id class_id

* Set seed and generate random number
set seed 123456
gen rand = runiform()

* Randomize the three classes within each school
bysort school_id (rand): gen assignment = _n

label define assignment_lbl 1 "Control" 2 "T1" 3 "T2"
label values assignment assignment_lbl

Here, school_id uniquely identifies each school and class_ID each class. After sorting classes by the randomly generated number within each school, assignment = 1 corresponds to Control, assignment = 2 to T1, and assignment = 3 to T2. The randomization seed, code, and resulting treatment assignment will be retained so that the procedure can be reproduced exactly.

School class

Yes

42 school classes

approximately 1,050–1,260 students

14 control classes, 14 treatment arm 1, 14 treatment arm 2

Pre-analysis plan