PATIENTS AND METHODS
Study design and patients
The CHECK patients used for this analysis were between ages 1-27y and had a Medicaid diagnosis of SCD. In some cases, SCD was among multiple chronic disease diagnoses (most commonly asthma, diabetes mellitus, or epilepsy).
Procedures
The CHECK program served families throughout the Chicago area. Initial eligibility was based on Medicaid claims data and is described extensively elsewhere 23. Patients were passively selected into the CHECK programs and CHECK program data were collected from January 1, 2015, through January 12, 2018. All patients were sent a letter stating that they were enrolled. CHECK CHWs attempted to contact a subset of patients (based on risk and diagnosis) either by mail, phone, or household visit31. Patients who completed the CHW-administered interview were considered as ‘engaged’ in the CHECK program. Those who did not complete the intake assessment were considered ‘enrolled’ but not ‘engaged’. Engaged patients were connected to CHWs who provided consultation, care coordination, education, and social support services as needed33. Patients were enrolled and participated in the CHECK program on a rolling basis over time. The CHECK data collection was approved by The University of Illinois at Chicago Institutional Review Board (protocol #2017-0604) 34.
Assessment and criteria
Inclusion criteria were Medicaid insurance, CHECK enrollment, and sickle cell disease diagnostic ICD9 or ICD10 code. Retrospective data for this study were extracted from Illinois Medicaid paid claims for a three-year period per participant: one year prior to CHECK enrollment (Baseline Year) and the following two years during CHECK enrollment. Exclusion criterion was diagnostic code for sickle cell trait. Based on Baseline Year Medicaid claims, patients were categorized as High, Medium, or Low risk for incurring inpatient expenditures during the CHECK enrollment period. High risk patients were those having more than 3 emergency department (ED) visits or were hospitalized more than once during the Baseline year. Medium risk patients were those who had 1 to 3 ED visits or 1 hospitalization during the Baseline year and Low risk patients were those who had no ED visits and no hospitalizations during the Baseline year33. Outliers with inpatient expenditures more than $100,000 per year in any CHECK year were excluded from analyses because such patients were expected to have unique medical problems beyond their SCD 5,33.
Statistical analysis
The analytic plan was developed to handle skewed data with outliers of high expenditures and individual heterogeneity over time, which are seen in all studies of SCD. Expenditure data were analyzed across three years for everyone, based on each individual’s enrollment in CHECK: a Baseline year preceding enrollment, then one year and two years after CHECK enrollment. Analyses were conducted using the R program, version 4.0. Outliers with inpatient expenditures more than $100,000 per year were already excluded from analyses 5,33. Because the overall distribution of expenditures was highly skewed, data were transformed by taking the natural logarithm of each patient’s expenditures to reduce the distortion caused by the high values. To account for the zero expenditure cases, the number one was added to every expenditure value to enable the logarithmic transformation. 11Geometric means were calculated as the nth root of the product of n logarithmically transformed expenditure values. They were used instead of arithmetic means because geometric means are appropriate summary statistics to report log-transformed data. It is the average of log-transformed value converted to the original expenditure scale. Geometric means of the log transformed data were calculated.
Baseline distributions of demographics and medical conditions were compared by enrollment risk using Pearson’s chi-square test or Fisher’s exact test. Analyses were conducted using the R program Version 0.7.15.
Because many SCD patients had no inpatient expenditures, a two-part expenditure analysis based on a statistical decomposition of the distribution of the outcome into a process that generates zeros and a process that generates non-zero positive values 35 was conducted using the GLMMadaptive (v0.7.15) R package. The analysis accommodates the semi-continuous expenditure data; that is, a continuous model allowing for data with excess zeros was fitted to the data36,37.
Using this approach, excess zeros were accounted for in an analytically appropriate way, so that better estimates of effects were produced. The model consisted of a logistic regression for the binary indicator that inpatient expenditures were zero or not and a standard linear mixed model for the log transformed non-zero inpatient expenditures. Interactions between utilization risk group and CHECK year were examined in both parts of the model. (Table 3). The first part of the analysis estimated the percent expenditure differences between utilization risk groups for each CHECK year while the second part estimated ratios of the odds of having zero expenditures between utilization groups for each CHECK year. For subgroup analysis, Wilcoxon pairwise tests were performed to compare mean inpatient expenditures over CHECK years (Baseline year, first year in CHECK, and second year in CHECK) within each utilization risk group. Multiple comparisons were accounted for by using the Bonferroni correction38.