Enabling Access to Free Higher Education in Bogotá through Personalized Admission Chances on Atenea (2026–2027)

We implement a personalized information-and-recommendation intervention integrated into Atenea's centralized higher-education assignment platform in Bogotá, Colombia, where applicants submit a rank-ordered list (ROL) of up to three programs and a placement bundles the seat with a full subsidy. The intervention has two components: (i) a risk widget shown on each program page that displays the applicant's estimated admission probability and waitlist/priority position; and (ii) a pre-submission popup that displays personalized admission probabilities for each listed option, message-flag warnings when the portfolio is risky (e.g., no safe option, over-concentration in highly competitive programs), and up to three recommended alternative programs with one-click links to add them. Applicants are randomized into three arms (Control / Widget / Widget+Popup). The goal is to improve the feasibility and quality of submitted lists, reduce non-assignment and improve college persistence at t+1.

Within the Widget+Popup arm a nested experiment varies how recommendations are generated and presented. The candidate pool in all sub-arms is the applicant's feasible set: programs with estimated admission probability above a 10% floor that are not already listed. Recommendation algorithm (1:1:1): (a) Random feasible-set, uniform draw from the feasible set; (b) Heuristic + 1 random, a content-similarity score weighting field/major-area similarity 0.5, admission probability 0.4, seat availability 0.1, returning the top-3, plus one random feasible program appended; (c) Model-based + 1 random, a latent-class (2-type) structural demand model estimated by EM that scores feasible programs by predicted contribution to expected portfolio value, plus one random feasible program. Ordering is independently randomized 50/50 (probability-ordered vs. random) to identify position/salience effects. The production widget classifies applicants into risk tiers (low >70%, medium 31–70%, high ≤30%) and escalates the popup's visual warnings as portfolio risk rises. Server-side assignment is via an independent uniform pseudo-random draw at first login (see Randomization Method).

2026-06-03

2026-06-10

Following the intervention's causal chain (behavioral response → ex-ante portfolio quality → realized assignment):
P1) Portfolio edited: indicator for any change (add/drop/replace/reorder) between the applicant's first recorded list and final submission (first-stage behavioral response).
P2) Δ Portfolio risk: change in probability of non-assignment, risk = ∏ₖ(1−pₖ); negative = safer portfolio.
P3) Δ Ex-ante probability of first-choice assignment: change in simulated probability of assignment to the top-ranked program.
P4) Assignment to first preference (realized): indicator for being assigned to the first-ranked option at the end of the cycle.

The two ex-ante quality outcomes (P2, P3) are measured as the change from the applicant's first recorded list to their final submitted list. The per-program admission probabilities pₖ are simulated from a structural model of the assignment mechanism (500-simulation bootstrap of cutoffs/quotas); portfolio risk is the product ∏ₖ(1−pₖ) over listed programs, and P3 is the simulated probability of being assigned to the top choice. Changes are reconstructed from the platform's immutable, append-only ROL snapshots (a new snapshot each time the list is updated); the first recorded list provides a common pre-intervention baseline for treated and control applicants alike. P4 comes from administrative assignment data (initial + waitlist rounds). Baselines (calibrated on the JE2 Bogotá eligible cohort): P1 editing rate 0.15 (assumed; not observable historically); P2 baseline risk level ≈ 0.90; P3 baseline ex-ante P(1st) ≈ 0.055; P4 first-preference rate 0.115.

Realized non-assignment and time-to-assignment (days, incl. waitlist dynamics); Δ ex-ante probability of assignment to any listed program and Δ expected rank of the assigned program; portfolio composition/safety (number of feasible options listed [adm. prob. > 10% floor], indicator for listing ≥1 such "safe" option, list size, distance to program, program type); adoption of recommendations (share of recommended programs in the final but not the pre-exposure list); assignment from waitlist vs. directly from the mechanism; program-level aggregates (mean priority index of offered students, rejection rate, waitlist cutoff, unfilled seats); and, among matriculants, dropout at t+1 and match-quality proxies (distance, field affinity, percentile relative to cutoff). Nested-experiment outcomes (within Widget+Popup, separate multiple-testing family): R1, recommendation adopted (indicator for adding ≥1 recommended program to the final list); R2, Δ Portfolio risk (as P2), compared across recommendation algorithms.

A "safe"/feasible option is a program with simulated admission probability above the 10% feasibility floor. Recommendation adoption (R1) is the share/indicator of recommended programs appearing in the final list but absent from the pre-exposure list. Program-level effects are constructed by aggregating applicant-level outcomes to the program × period level, using the treated share of a program's applicants as treatment intensity. Persistence/match outcomes use administrative data tracking applicants into higher education at t+1.

A three-arm individual-level randomized controlled trial on Atenea's centralized assignment platform in Bogotá, with 1:1:1 allocation to Control, Widget, and Widget+Popup. Randomization is server-side with rolling (just-in-time) assignment the first time an applicant logs in. The primary estimand is the Intention-to-Treat (ITT) effect, estimated by OLS of each outcome on arm indicators with baseline covariates (ICFES/Saber-11 score and sex) for precision and HC1 robust standard errors; family-wise error across the four primary outcomes is controlled via Holm–Bonferroni. A nested experiment within the Widget+Popup arm independently randomizes the recommendation algorithm (3 sub-arms) and presentation order (2 sub-arms).

ITT specification: Yᵢ = α + Σₐ βₐ·1{Tᵢ=a} + γ′Xᵢ + εᵢ (no stratum fixed effects, since randomization is simple/unstratified; Xᵢ = pre-intervention list composition, Saber-11, sex). TOT/LATE: exposure/use (popup seen, clicks, adoption) instrumented by assigned treatment in 2SLS; first-stage F and complier characteristics reported. Nested contrasts: among treated, effects of recommendation algorithm (model vs. heuristic vs. random) and order (probability vs. random) on adoption (R1), portfolio risk (R2), and assignment, with a separate Bonferroni family. Program-level/spillovers: difference-in-differences using a program's treated-applicant share as intensity, with program and period fixed effects. Heterogeneity: deciles of baseline priority index, municipality, gender, digital-access proxies, and administrative SES. Data: platform interaction logs + append-only ROL snapshots (P1–P3), Atenea administrative assignment/waitlist/enrollment data (P4, secondary), supply-side capacity/quota/cutoff data, and a post-application/pre-results survey (beliefs, comprehension, true preferences via best-worst scaling, satisfaction).

Randomization done by computer. Server-side randomization module assigns each applicant on first login via an independent uniform pseudo-random draw, with equal (1/3) probability across the three arms (rolling/just-in-time assignment; assignment is persisted and never re-drawn). Within the Widget+Popup arm, the recommendation algorithm is assigned by an independent uniform 1/3 draw and the presentation order by an independent 50/50 draw.

Individual applicant. (Two nested sub-randomizations — recommendation algorithm and presentation order — are also at the individual level, within the Widget+Popup arm.)

No

Not applicable. individual-level randomization (no clusters). Equivalently, the unit count equals the number of individually-randomized applicants (≈ 32,000; see below).

≈ 32,000 applicants (power calibrated on the JE4 Bogotá applicant cohort, N = 32,306). Enrollment is rolling over the application cycle

Three arms, equal allocation: ≈ 10,769 Control / ≈ 10,769 Widget / ≈ 10,768 Widget+Popup (Monte Carlo balance design, N = 32,306). Nested experiment within Widget+Popup (n ≈ 10,800): recommendation algorithm ≈ 3,590 per sub-arm (random / heuristic / model); presentation order ≈ 5,385 per sub-arm (probability-ordered / random).

Method: 500 Monte Carlo replications per effect-size grid point; outcomes simulated under JE2-calibrated baselines and re-estimated with the ITG specification (ICFES + gender controls, HC1 robust SEs, no clustering and individual randomization). Family-wise error across the four primary outcomes controlled via Holm–Bonferroni; MDE = smallest effect reaching 80% adjusted power at α = 0.05, with N ≈ 32,000 (~10,700/arm). Primary outcomes (Holm–Bonferroni-adjusted, 80% power): | Outcome | Baseline | SD | MDE | Normalized | |-----------------------------|---------------------|--------|----------|---------------------| | P1 Portfolio edited | 0.15 | — | +1.8 pp | 12% of baseline. | | P2 Δ Portfolio risk | 0.0 (level ≈0.90) | 0.27 | −0.012 | 0.04 SD | | P3 Δ Ex-ante P(1st) | 0.0 (level ≈0.055) | 0.20 | +0.012 | 0.06 SD | | P4 Assigned to 1st pref | 0.115 | 0.319 | +1.8 pp | 16% of baseline | (SD for P2/P3 is the cross-sectional SD of the level, a conservative upper bound on the SD of the change, so true MDEs are likely smaller.~1.8 pp on P4 ≈ 190 additional students placed in their top choice.) Nested recommendation experiment (pairwise, 80% power, Bonferroni within nested family); ρ = report-engagement rate: | Outcome | Algorithm contrast (n≈3,590/arm) | Ordering contrast (n≈5,385/arm) | |--------------------------------------------------|---------------------------------------|--------------------------------------| | R1 Recommendation adopted (base 0.15) | 2.8 pp (ρ=1) / 3.6 pp (ρ=0.6) | 2.3 pp / 3.0 pp | | R2 Δ Portfolio risk (SD 0.27) | 0.021 / 0.027 | 0.017 / 0.022 | | Δ Ex-ante P(1st) (SD 0.20) | 0.016 / 0.020 | 0.013 / 0.017 | The study is well powered for the four primary ITT outcomes and for nested recommendation adoption (R1); the nested portfolio-risk contrast (R2, MDE ≈ 0.02) resolves only sizable algorithm differences, and the algorithm×ordering interaction (six cells ≈ 1,800 each) is underpowered and treated as exploratory.