Search
Browse By Day
Browse By Time
Browse By Person
Browse By Policy Area
Browse By Session Type
Browse By Keyword
Browse Artificial Intelligence Presentations
Program Calendar
Sign In
Search Tips
Background: Childcare subsidy receipt is shaped by a complex set of family eligibility, provider participation, policy changes, and local childcare market conditions. Yet state agencies often lack small-area forecasting tools to identify where subsidy receipt is likely to rise or fall over time. Evidence exists on descriptive, individual-level patterns of subsidy use, yet offering limited insight into geographic and structural factors shaping utilization. Because subsidy receipts likely reflect nonlinear interactions among potential predictors, conventional approaches may not fully capture these dynamics. This study examines whether linked administrative data and machine learning can improve predictions of geographic variation in subsidy receipt across Massachusetts and identify factors most strongly associated with predictive performance.
Methods: We used Massachusetts administrative data from 2016 to 2024 to build a census tract-by-month panel dataset by linking subsidy records, provider licensing data, tract-level demographic indicators, and policy timing measures to predict the monthly number of children receiving subsidies in each census tract (N=148,632 tract-month observations). Predictors included provider-market characteristics, licensed childcare capacity, subsidy density, local demographics, seasonality, and policy shifts. We estimated fixed-effects models using the full sample as the baseline and a temporal holdout split (training 2016-2022; test 2023+), comparing a mean year-month FE baseline and a linear time trend specification as a robustness check. An XGBoost model was estimated using a 10% subsample split into training (80%) and test (20%) sets. Model performance was evaluated using root mean squared error (RMSE), mean absolute error (MAE), and R-squared. Feature importance and SHAP analyses were used to interpret the XGBoost results. Results: The baseline fixed-effects model achieved a within-R2 of 0.805 in-sample (RMSE = 6.61, MAE = 3.01). Out-of-sample performance on the 2023+ holdout was consistent across both prediction strategies: mean year-month FE yielded (RMSE = 7.54, MAE = 3.41, R2 = 0.986) and the linear time trend yielded (RMSE = 7.55, MAE = 3.54, R2 = 0.986). Prior-month subsidy count and local provider-market conditions were the strongest predictors. The XGBoost model showed strong predictive performance in forecasting local monthly subsidy receipt with high precision (RMSE = 4.31; MAE = 1.66; R2 = 0.995). The strongest predictors were provider-market conditions, including licensed childcare capacity, concentration of subsidized children, lagged licensed capacity, preschool capacity, and school-age capacity. Even after removing several dominant proxy-like predictors, model performance remained, suggesting that prediction was driven not only by near-outcome indicators but also by broader structural features of local childcare markets.
Implications: Findings highlight the promise of linked administrative data and machine learning for improving local prediction of subsidy receipt and developing actionable forecasting tools for subsidy systems. Results also suggest that subsidy receipt is shaped not only by family demand, but also by local childcare market structure and capacity. Forecasting tools may help state agencies proactively identify emerging access gaps, anticipate changes in local demand, and target reimbursement and provider support to strengthen equitable access to subsidized care. Future analyses will implement directly comparable approaches to assess the relative performance of fixed-effects and machine learning models.