Who Returns to SYEP? Predictors of Youth Program Reapplication

Overview

Analyzed which participant, placement, and provider characteristics predict whether young people reapply to NYC’s Summer Youth Employment Program (SYEP). I did this as an analytics intern at the NYC Department of Youth and Community Development (DYCD), Bureau of Planning, Program Integration, and Evaluation, in Summer 2026. The analysis covers Older Youth, Community-Based participants in 2022–2024 (130,417 person-years) and asks who applies again the following year. Reapplication, rather than re-enrollment, is the outcome because re-enrollment is capped by program capacity: it would measure the lottery as much as the participant.

GitHub

Technical Approach

Data & Frames

I built the analysis from raw application extracts (~1M rows, 2019–2025) and pre/post participant surveys (~151K rows). That took three main pieces of data work:

• Identity resolution. The system issues IDs per application, not per person, so I linked applications across years through an agency-provided participant-ID crosswalk.
• Cross-year worksite matching. Worksite IDs are reassigned every cycle, so I built a fuzzy-matched bridge to recover persistent sites.
• Codebook-based cleaning. Free-text fields like borough were mapped to a codebook, with a manual review step for uncertain matches.

The work produced two deliberately separate frames:

• Frame A (administrative): 130,417 person-years. This frame carries the headline estimates.
• Frame B (survey): 5,844 participants, one row each, taken at their first observed year. Survey response requires enrollment, so respondents are a selected group. Frame B is therefore never pooled with Frame A.

Models

Logistic regression with standard errors clustered by provider. Predictors:

• an age spline
• six hours-worked bands, with full-time work (126–149 hours) as the reference
• year and borough
• placement-quality flags
• race and gender

Four fits isolate different questions:

• Headline model on the full administrative frame
• Person-clustered refit as a dependence check
• Bridge model that runs the administrative specification on survey respondents only, separating who answered from what they answered
• Survey-augmented model with 12 pre-specified items

Benjamini–Hochberg correction was applied within each fit. Secondary analyses:

• g-computation by age group
• block-level Wald tests
• a provider analysis that estimates each provider’s return rate relative to its participant mix

Results

How much of the program a young person completes is by far the strongest predictor of return. The relationship is monotonic, with no inversions across six hours bands:

What participants said about the program predicted almost nothing:

• Benefits were precise nulls. None of the five self-reported benefit items survived correction, and effects larger than ~1.2× were ruled out.
• One survey item did predict return: whether SYEP was their first choice (OR 1.32, q = 0.003).

Together, these point to allocation and completion, not post-program satisfaction, as the places where retention is determined.

Providers differ significantly in adjusted return rates (roughly −9 to +10 pp). Participant mix explains only ~20% of that spread.

Reflections

The biggest lesson was how much depends on inference and selection decisions:

• Clustering changed a conclusion. Borough was strongly significant under person-clustered standard errors but null under provider clustering. Providers nest within boroughs, so the provider-clustered result is the honest one.
• The survey sample is a completer population. Respondents return at 69.6% vs. 56.1% for non-respondents, and only 1.2% worked zero hours (vs. 20.0%). Survey findings describe completers, not all participants.
• These are associations, not causal effects. Hours worked is endogenous: the engagement that keeps someone on site likely also brings them back.

Because this is real agency data, I also built a de-identification step for the public repository. Provider organizations get random pseudonyms, banded sizes, and rounded rates, so no provider can be identified.

Future directions:

• Use the model to determine more abouts early attrition, where it is most actionable.
• Advocate for a worksite ID that persists across cycles.
• Test whether providers’ relative positions are stable year to year before using them for any decision.