Statistical Analysis
We first characterized data missingness and calculated descriptive statistics for the sample. Patients without a valid street addresses and those with missing clinical data were excluded from the analytical sample. As less than 7% of the sample had missing clinical data, we used listwise deletion to account for missingness. A STROBE diagram of the study population is presented in Figure 1 .
The distribution of neighborhood-level and patient characteristics was obtained for the overall sample and by recurrent ACS status. Bivariate associations were estimated using simple logistic regression. Models of recurrent ACS were estimated with multiple logistic regression. As 52% of Census block groups included singleton observations with only one individual per block group, we used robust standard errors to account for clustering of individuals within block groups.21We estimated generalized linear models to compare recurrent ACS to single and no ACS combined, with results represented as incidence-rate ratios (IRR). We estimated multinomial logistic regression models to compare each category (no ACS/single ACS/recurrent ACS), with results represented as relative risk ratios (RRR). To better understand the contribution of residential characteristics to ACS recurrence, we four separate models were estimated: Model 1 with neighborhood socioeconomic deprivation alone, Model 2 with neighborhood racial composition alone, Model 3 with rurality alone, and Model 4 with all three characteristics together. All four models included the same covariates: age, sex, insurance type, BMI, chronic transfusions, hydroxyurea, asthma, and phenotype. Statistical tests were two-sided and were performed using a 5% significance level (α=0.05). Analyses were performed using Stata software, version 15 (StataCorp, College Station, TX).