Setting and Study Population
We conducted this study within two integrated care delivery systems with membership representative of the larger communities they serve—Kaiser Permanente Northern California (KPNC) and Kaiser Permanente Colorado (KPCO).16 Eligible study members included adults aged 18 years or older on March 1, 2020, who completed a Medicaid new member survey within the past 2 years (KPNC) or an annual Medicare Health Risk Assessment survey (KPCO) within the past year. These member surveys are administered by health system staff and documented in the member’s electronic health record captured member-identified health status, living situation, and social risk factors including food insecurity, financial strain, and housing concerns.
We measured SARS-COV-2 testing and positivity rates between March 1, 2020, and November 30, 2020. Member survey responses, demographics, diagnoses, and laboratory results were extracted directly from the KPNC and KPCO EPIC-based electronic medical record systems and from each site’s Virtual Data Warehouse, a research database that aggregates member health data in a consistent, standardized format from multiple internal sources and externally billed claims.17
Exposure and Covariates
The primary exposure was self-reported race/ethnicity. Member race/ethnicity and need for interpreter services were captured at the time of health plan enrollment or by medical assistants and clerks during outpatient or inpatient care contacts. We determined chronic medical conditions using the ICD-10 diagnosis code list from the Centers for Medicare and Medicaid Chronic Conditions Warehouse, and we calculated the Elixhauser comorbidity score using 2 years of historical visit information.18 We defined members as obese based on body mass index > 30 kg/m2 as calculated from most recent visit height and weight measurements. Social factors were coded from responses to member health assessment questions about worry related to food running out; eating fewer than two meals a day; trouble paying for basics such as food, housing, and utilities; concerns about housing safety, affordability, or transience; lack of transportation to make medical appointments or do activities of daily living (ADLs; e.g., getting out of bed or chair, bathing/showering/dressing/eating); and limitations on performing ADLs and instrumental ADLs (IADLs; e.g., managing medicines and finances, shopping for groceries). Responses also provided information about the presence of others in the household or if the member did not live independently.
The U.S. Census–based Social Vulnerability Index (SVI) score and component measures were obtained from the CDC and linked to member geocoded addresses.19 The SVI composite sum score ranges from 0 to 15 and incorporates measures of census tract demographics, socioeconomic status, and housing factors such as unemployment, poverty, and proportion of residents living in crowded housing.
Our primary outcomes were SARS-CoV-2 testing and infection between March 1 and November 30, 2020. We designated members as tested if their electronic medical records contained a completed SARS-CoV-2 PCR test or outpatient and inpatient visit with a COVID-19 screening ICD-10 diagnosis. Among the tested members, we classified those with a positive result for SARS-CoV-2 laboratory test or a confirmed COVID-19 diagnosis as positive for COVID-19 infection.
We used a staged modeling approach to describe differences in testing and positivity by race/ethnicity, and we examined the contribution of measured social risk factors to these differences. We assessed the baseline cohort characteristics on March 1, 2020.
We performed logistic regression on the dichotomous outcomes of having been tested (entire cohort) and of testing positive (among those who were tested). To obtain risk estimates, we fit a population-average model with generalized estimating equations and an independent correlation structure. This approach appropriately accounted for the within-cluster correlation in covariates measured at the census tract-level (Social Vulnerability Index score) among members residing within the same census tract and is robust to misspecification of the working correlation structure.20 Model 1 is adjusted for age, sex, and need for an interpreter. Model 2 represents added medical and mental health conditions that are risk factors for poor COVID-19 outcomes or are potential confounders due to their influence on the likelihood a medical provider would order testing and the members’ ability to obtain testing.25 Conditions included diabetes, chronic obstructive pulmonary disease, asthma, hypertension, cardiovascular disease, hyperlipidemia, depression, anxiety, Alzheimer’s disease, and related dementias, and the Elixhauser score for a measure of overall comorbidity burden. The final model, Model 3, represents added dichotomous self-reported social factors and the continuous neighborhood-level SVI score. Social factors in the KPNC models included living with others, children in household, education less than high school, trouble paying for basics, food insecurity, housing concerns, and help needed with ADLs. Social factors in the KPCO models included education less than high school, not living independently, food insecurity, and limitations in ADLs and limitations in instrumental ADLs.26,27 All descriptive and statistical analyses were performed using SAS 9.4. Our research was approved by the Kaiser Permanente Institutional Review Board and all procedures followed were in accordance with the ethical standards of the IRB and the Helsinki Declaration of 1975, as revised in 2000.