Methods/Statistics
Reassessing positive results: Uncovering potential detection bias in epidemiologic studies of dementia using electronic health record data Jingxuan Wang* Jingxuan Wang Minhyuk Choi Scott C Zimmerman Ruijia Chen Catherine Schaefer Deborah Blacker M. Maria Glymour
Many epidemiologic studies of dementia use electronic health record (EHR) data, which are vulnerable to detection bias. Dementia is more likely to be detected when patients have more contact with the healthcare system, which can lead to a systematic difference in the ascertainment of outcomes across exposure groups. Using EHR data from the UK Biobank, we evaluated potential detection bias by estimating associations of two negative control exposures with dementia.
UK Biobank participants aged ≥ 55 without baseline dementia (N=141,060; mean age=62.5 years) were followed for a median of 13.7 years. We estimated associations of forearm fracture and urinary tract infection (UTI) with incident dementia. Diagnoses were from primary care, hospital inpatient, and death record data. We used Cox models to evaluate associations of binary exposure with all-cause dementia overall and in 4 time intervals after the exposure (0–1, >1–5, >5–10, and >10 years) compared to individuals with no prior exposure. All models adjusted for age, sex, race, education, assessment center, and healthcare utilization (number of clinic visits during the year before baseline).
Prior to baseline, 6,843 (4.9%) and 16,742 (11.9%) individuals had a forearm fracture and UTI diagnosis; 3,896 dementia cases accrued during follow-up. Forearm fracture (HR=1.19, 95%CI=1.06-1.33) and UTI (HR=1.83, 95%CI=1.70-1.96) were associated with dementia incidence, especially in the first year after exposure diagnosis (forearm fracture HR=1.93; UTI HR=6.53). The association attenuated in subsequent years (Figure).
The chance of receiving a dementia diagnosis increases sharply in the first year following a forearm fracture or UTI diagnosis, although the influence of either condition on dementia pathology is likely small. These short-term increases likely represent detection bias, i.e., they would not have received a dementia diagnosis at that time in the absence of the medical attention necessitated by forearm fracture or UTI.
