Level-Up: Understanding Parental Roles in the Career Choices of High-School Graduates

We study a 2×2 randomized intervention in which households (high-school students matched to a parent or guardian) are given access to a multilingual AI chatbot delivered through a study smartphone app. All households receive the chatbot. A specialized financial-aid module is activated for the student, the parent, both, or neither, yielding four treatment arms.

2026-06-01

2027-10-31

1. Realized post-high-school pathway

2. Path–preference match

3. Field quality

4. Distance from parental home

5. Corrected financial-aid beliefs

6. Financial-aid application

1. Realized post-high-school pathway (categorical): university enrollment, vocational training (company-based BAB or school-based BAföG), gap year, direct labor market entry, other. This is measured in the endline.

2. Path–preference match (binary): two binary indicators capturing whether the realized pathway matches (a) the student's own unconstrained pathway preference and (b) the parental unconstrained pathway preference, both elicited at baseline (and updated at midline/endline). This allows us to distinguish four constellations: match with both, match with own preference only, match with parental preference only, and match with neither. For deviations from the student's own preference, we further distinguish between forced and willing deviation, i.e. applied and rejected vs. never applied (based on endline information).

3. Field quality (continuous, composite): selectivity (Numerus Clausus cut-off, where applicable) and institution ranking in usual field ranking, combined into a standardized index.

4. Distance from parental home (continuous, km): geodesic distance from the parental residence to the chosen institution or training employer. In case of missing parental residence, we assume the school to proxy parental residence.

5. Corrected financial-aid beliefs: absolute deviation of self-reported beliefs about (a) own/child's eligibility (binary), (b) expected monthly aid amount (continuous), (c) amount difference when living at home (continuous) and (d) repayment conditions from the true value implied by the participant's household characteristics (continuous).

6. Financial-aid application (binary): indicator equal to 1 if the student filed or is intending to file an application for BAföG or BAB by the endline survey.

1. Chatbot information demand

2. Intra-household decision weights

3. Heterogeneity

4. Intra-household communication

1. Chatbot information demand: number of financial-aid-related queries, time-to-first-financial-aid query, return-visit count, total session time on financial-aid topics, and the breakdown of queries by topic (eligibility, amount, repayment, application).

2. Intra-household decision weights: structural estimates of a Pareto weight in a collective household model with subjective beliefs about aid amounts. This is conditional on seeing a successful first stage effect that our treatment changes beliefs.

3. Heterogeneity: treatment effects by socioeconomic status, first-generation-college status, migration background, baseline misperception severity, east-west and gender.

4. Intra-household communication: self-reported frequency and timing of parent-student conversations about post-school plans, including timing relative to chatbot interactions.

2×2 cluster-randomized field experiment with students in their penultimate year of German high school (Abitur) and their parents/guardians. Participants register through a study smartphone app and are followed for 17 months across baseline, midline, and endline surveys plus monthly pulse measurements. Randomization to the four arms is at the school level. The primary analysis is intention-to-treat with standard errors clustered at the school level.

Not available

The chatbot-arm randomization is performed ex ante at the regional school-cluster level over the full population of eligible schools in Germany. Each school ID is mapped to a school site, and sites are grouped into regional clusters based primarily on geographic proximity, using coordinate-based grid cells where latitude/longitude are available, with a postal-code fallback where coordinates are missing. Duplicate or same-location school IDs are grouped before clustering. Randomization is stratified by federal state. Within each state, regional clusters are assigned to the four arms of the 2x2 design (control, student-only access, parent-only access, and access for both student and parent). Rather than an independent uniform draw per cluster, the assignment is optimized to balance the arms within state, with highest priority on the number of clearly Abitur-granting schools and the number of already-registered student/parent-child pairs, and lower priority on likely-Abitur schools and unclear school-like cases. The assigned arm is then mapped back to every school ID in the cluster, so that all schools in the same regional cluster receive the same treatment arm. Each participant from a given school, whether already registered or registering later, inherits that school's arm with no further draw. Realized balance on the stratification and clustering covariates, and on additional covariates including urban vs. rural location and school-mean socioeconomic status, will be reported, and these variables will be included as controls in ITT specifications for robustness.

The "Level-Up" module-order randomization for the three gamified educational modules following the initial fake-news module is performed at the individual student level using a uniform random draw over the six possible orderings, independent of chatbot assignment.

The chatbot intervention is randomized at the regional school-cluster level. Geographically proximate schools are grouped into clusters, each cluster is assigned to one of the four arms, and the assignment is mapped back to the school IDs within the cluster so that all participants from a given school receive that school's (cluster's) assigned arm.

The module-order randomization is at the individual student level and is orthogonal to the chatbot treatment assignment.

Yes

The randomization unit is the regional school-cluster, so the relevant cluster count is the number of regional clusters formed across the full population of eligible schools in Germany, not the number of individual schools. Germany has approximately 3,100 Gymnasien at the Abitur level. Grouping geographically proximate schools, and merging duplicate or same-location school IDs, reduces this to a smaller number of regional clusters, which sets the upper bound on the number of randomization units. The number of clusters and schools actually represented in the analysis sample is determined ex post by the set of schools among realized participants, as recruitment is conducted top-down via app sign-ups rather than through pre-recruited schools. Based on realistic uptake assumptions from our recruitment channels, we expect several hundred schools to be represented, corresponding to a somewhat smaller number of regional clusters, with the precise figures depending on sign-up dynamics over the recruitment window.
To improve geographic and demographic coverage, we reserve the right to conduct targeted recruitment campaigns in regions or school types that are underrepresented relative to the national distribution at interim monitoring points. Chatbot-arm assignment is fixed at the cluster level for the full population of eligible schools, so such targeted outreach affects only which pre-randomized clusters and schools become active and does not change the arm assigned to any given school. Any targeted campaigns conducted will be documented and reported.

19,612 students, each matched to at least one parent or guardian, for a total of approximately 39,000 individual participants. Minimum sample we aim for: 10,000 students (≈20,000 individual participants once parents are matched), corresponding to the floor scenario for our recruitment outreach.

We target an approximately equal split across the four arms, around one quarter of clusters per arm, within each federal state. Because clusters are assigned through a balancing procedure at the regional school-cluster level rather than an independent uniform draw, cluster counts per arm are close to equal but need not be exactly equal, since the procedure prioritizes balance on the key covariates (the number of clearly Abitur-granting schools and already-registered student/parent-child pairs first, likely-Abitur and unclear school-like cases second). Further deviations from exact equality are expected because (a) sign-ups remain open until the summer break and (b) realized cluster sizes vary across schools. Realized arm sizes will be reported, both in terms of clusters and schools and in terms of participants.

Accounting for school-level clustering, our power analysis indicates that 19,612 students are required to detect a 2 percentage-point change in a binary primary outcome (e.g., financial-aid application, university enrollment) at 80% power and a 5% significance level, under design-realistic assumptions about the school-level intra-cluster correlation. At our floor sample of 10,000 students, the minimum detectable effect on the same outcome rises to approximately 2.8 percentage points under the same power and significance level.