Evaluating the Impact of Data Standardization on Administrative Healthcare Data Among Patients with COVID-19 Diagnosed in Outpatient Settings

Background: Common data models apply a standard format that attempts to mitigate variation across disparate real-world data sources, but the impact of the transformation on data completeness and subsequent validity of study findings remains unclear.

Objectives: Describe differences in key data elements important in COVID-19 research using administrative healthcare data after applying the Sentinel common data model (SCDM) to native formats.

Methods: Patients were selected from the same HealthVerity medical and pharmacy claims, hospital chargemaster, and external commercial laboratory data in untransformed (native) and in SCDM-transformed formats. The cohort comprised all patients with an outpatient COVID-19 diagnosis claim between 01-Jan-2021 to 31-Dec-2022. Cohort attrition and sample size, patient characteristics, outpatient COVID-19 treatments, healthcare utilization (HCU) (e.g., counts of different visit types) during 180-day baseline period and incidence of medical conditions of interest after COVID-19 diagnosis were described for the native and SCDM formats.

Results: Fewer patients were identified using the SCDM-transformed versus native data (164,445 vs. 198,317). Patients in the SCDM-transformed data had fewer orders for any SARS-CoV-2 diagnostic tests (59% vs. 70%), a lower incidence per 1000 person-years of health conditions of interest (ranging from 8.8 vs. 9.7 for hypersensitivity to 29.6 vs. 37.1 for hepatotoxicity), and less HCU. For HCU counts, far fewer events were identified in the SCDM-transformed data that collapses multiple events per day than the native data. For example, mean baseline specialty care observations were 6.0 SCDM vs. 27.7 native, raw count observations with no restrictions imposed; differences were mitigated when counts were restricted to one visit per day (e.g., 2.9 vs. 3.5). The distribution of patients with dispensing claims for COVID-19 treatments were similar. In post-hoc analyses that manipulated the native cohorts to include the same patients selected via the transformed data, differences remained in the distribution of key variables between transformed and native data.

Conclusions: In this application of SCDM transformation, dispensing events were similar as most pharmacy attributes were retained. However, laboratory results, HCU, and medical conditions had lower capture than in the native format. Although there are benefits to the SCDM transformation collapsing redundant information (e.g., avoid overcounting of HCU), findings highlight some potential loss in data that should be further explored.