Human Capital as Insurance: Skill Portfolio Choices under Occupational Uncertainty

The experiment follows a 2×2 factorial design. Both information treatments are randomized at the classroom level and embedded within a baseline survey.

Treatment 1 — Labor Market Information (LM): Track-specific information about the distribution of occupational destinations among recent graduates. The information is drawn from administrative matches of prior cohorts (RAIS + SENAI) and shows that graduates disperse across a range of occupations rather than entering a single expected occupation. For each track, the module displays one expected occupation, three alternative occupations, and a residual category, along with track-level facts on vacancies, wages, and projected trends. Students see a neutral transition screen in lieu of SS.

Treatment 2 — Self-Information (SS): Personalized feedback on the student's relative standing in three soft skill domains (collaboration/teamwork, planning/organization, self-regulation), based on a validated baseline assessment battery. Feedback is presented as a percentile or coarse category relative to peers, highlighting strengths and areas for improvement. Students see a neutral transition screen in lieu of LM.

Combined arm (LM + SS): Both modules presented sequentially—labor market information first, then self-information.

Control: No information screens; students see a neutral transition screen of similar length.

All students (including control) complete the same baseline assessment modules before any information is shown, and the same outcome modules after.

2026-03-16

2026-12-31

1. Beliefs about occupational outcomes (perceived occupational mismatch)
2. Hypothetical portfolio allocation toward transferable skills
3. Stated willingness to enroll in a future soft skills support program

1. **Perceived occupational mismatch**: Students report a probability distribution over possible occupational destinations before and after treatment. The primary measure is the perceived probability of working outside the expected occupation group. We examine both the post-treatment level and the change from baseline.

2. **Portfolio allocation**: Students allocate a fixed time budget across three skill categories (track-specific technical, general technical, and soft skills). The primary outcome is the share allocated to transferable skills (general technical + soft skills).

3. **Stated take-up**: Binary measure of whether the student expresses willingness to be contacted for a future soft skills support program.

1. Alternative measures of perceived occupational dispersion
2. Perceived usefulness of soft skills (in expected occupation and across occupations)
3. Soft skill self-assessment beliefs
4. Wage beliefs and expectations
5. Labor market outlook and job search intentions
6. Follow-up behavioral outcomes (contingent on program implementation)
7. Longer-term labor market outcomes (contingent on data access)

1. **Occupational dispersion**: Alternative summary measures of the perceived occupational distribution (e.g., entropy, concentration indices) to complement the primary mismatch measure.

2. **Perceived skill usefulness**: Whether students perceive soft skills as useful in their expected occupation and as transferable across occupations. We also construct a "transferability wedge" capturing the perceived relative advantage of soft skills across vs. within occupations.

3. **Soft skill self-assessment**: Self-ratings and comparative assessments of own soft skills in three domains (collaboration/teamwork, planning/organization, self-regulation), measured before and after treatment.

4. **Wage beliefs**: Expected wages in the expected and non-expected occupations, and subjective probabilities of meeting various wage thresholds.

5. **Labor market outlook and job search**: Career predictability, job probability beliefs, reservation wage, preparedness for alternative occupations, job search strategies, and expected search intensity.

6. **Follow-up behavioral outcomes (contingent)**: If a post-baseline soft skills support program is implemented, we will measure engagement from administrative data. These outcomes are contingent on program launch.

7. **Longer-term labor market outcomes (contingent)**: Contingent on follow-up data access, we plan to measure employment status and formality, job quality (wages, contract type), skill utilization in the workplace, and job satisfaction. We also plan to link respondents to the RAIS formal-sector administrative dataset and to SENAI's graduate survey (Pesquisa de Egressos) to track longer-run labor market trajectories.

We implement a 2×2 factorial experiment embedded within a baseline survey administered to approximately 34,000 students across roughly 100 schools, 1,200 classrooms, and 47 vocational tracks in São Paulo, operated by SENAI-SP.

Both treatments are randomized at the **classroom level** to limit within-class contamination and spillovers during survey administration.

Classrooms are assigned to one of four strata groups before randomization, based on data availability:

1. Standard tracks (≥10 classes per track): Each track forms its own stratum; classrooms randomized into all four arms. Combined with small tracks, this group comprises 1,092 classrooms (~30,600 students), with 273 classrooms per arm.
2. Small tracks (<10 classes per track): Pooled into sector-level strata when the sector contains at least 10 classes; randomized into all four arms. Included in the 1,092 classroom count above.
3. Early-survey classrooms (5 schools, surveyed before soft skill percentile scores are available): 61 classrooms (~1,700 students). Cannot receive the SS treatment; randomized into LM or Control only (30 control, 31 LM).
4. No RAIS match classrooms (11 tracks without RAIS–SENAI administrative match data): 46 classrooms (~1,300 students). Cannot receive the LM treatment; randomized into SS or Control only (23 per arm).
5. Both constraints (early-survey and no RAIS match): 2 classrooms (~55 students). Assigned to Control only.

This yields unequal arm sizes: Control 328 classrooms (~9,200 students), LM only 304 (~8,500), SS only 296 (~8,300), LM+SS 273 (~7,600). Total: 1,200 classrooms, approximately 34,000 students.

All estimation uses classroom-level clustered standard errors and includes strata fixed effects.

Not available

Randomization done by a computer

Classroom (turma). Both Treatment 1 (labor market information) and Treatment 2 (self-information) are randomized at the classroom level. Randomization is stratified by school and course track, with strata defined as described above (standard tracks, small tracks pooled by sector, early-survey classrooms, and no-RAIS-match classrooms).

Yes

1201 classes

~34,000 students

- Control: 328 classrooms, ~9,200 students
- LM only (labor market information): 304 classrooms, ~8,500 students
- SS only (self-information): 296 classrooms, ~8,300 students
- LM + SS (both): 273 classrooms, ~7,600 students

Note: Unequal arm sizes arise from restricted-eligibility strata. Early-survey classrooms (61) can only be assigned to LM or Control. No-RAIS-match classrooms (46) can only be assigned to SS or Control.