Experimental Design
To construct our study sample, we developed a predictive model to estimate the likelihood each child not receiving SSI is in fact eligible, and then applied it to all children in Allegheny County who were under 16 years old as of April 30, 2026 and were receiving Medicaid but not SSI as of this date. 67,555 children met these basic characteristics. We used the fitted values of the model to assign each child an SSI risk score.
We further restricted the risk-scored pool in two ways. First, we removed children who were enrolled in Medicaid under the "PH95" eligibility category. This category covers children with disabilities whose parents earn too much money to qualify for traditional income-based Medicaid. After removing these children, we are left with 60,861 children in the risk-scored pool. Second, we included only one child per household. This was intended to prevent treatment spillovers within households. We kept the child in each household who has the highest SSI risk score. This led us to drop 25,861 children, reducing the pool to exactly 35,000 children.
For the final study sample, we selected the 3,500 children in the remaining risk-scored pool who were in the top decile of risk scores. These top-decile children have much more intensive usage of health care than the children who have lower risk scores. In the year before April 30, 2026, children in the top decile have more than eight times as many days with a mental health care Medicaid claim, eight times as many days with a claim for an inpatient hospitalization starting in the emergency room (ER), and nearly five times as many days with a claim for outpatient care not in the ER than the lower-risk score children. Children in the top decile also have six times as many distinct mental health diagnoses, with particularly higher rates of autism spectrum disorders and intellectual disabilities. Demographically, children in the top decile are somewhat more likely to be older, non-Black, and male, and to come from households with higher earnings, than the lower-risk score children.
We will then randomly assign these 3,500 children to one of the three study arms, as highlighted in the intervention design. For our primary analysis, we will use the random assignment to estimate impacts in a simple intent-to-treat framework. We will make the following comparisons to test the effectiveness of our intervention: (1) Information-only vs. control; (2) Information-plus-assistance vs. control; (3) Information-plus-assistance vs. information-only. If the impact of the information alone is sufficiently small or insignificant, we will also estimate impacts pooling the information-only and control groups together.
For more complete details, please see the attached full pre-analysis plan.