Robustly measuring multimorbidity using disparate linked datasets.

Background: Measurement of multimorbidity, the co-occurrence of two or more conditions in the same individual, is highly variable which limits the consistency and reproducibility of research.

Methods: Using data from 172,563 UK Biobank (UKB) participants and a cross-sectional approach, we examined how choice of data source affected estimated prevalence of 80 individual long-term conditions (LTCs) and multimorbidity. We developed code-list-based algorithms to determine the prevalence of 80 LTCs in (1) primary care records, (2) UKB baseline assessment, (3) hospital/cancer registry records, and (4) all three data sources together.

Results: Using records from all three data sources, 146,811 (85.1%) participants have at least one and 109,609 (63.5%) have at least two LTCs at baseline. A median of 4.7% (IQR 1.0-16.6) of participants with a condition are identified by all three data sources. Agreement is highest for endocrine, nutritional and metabolic disorders, with a median of 32.9% (IQR 20.5-34.1) of individuals with a condition identified by all three data sources. Agreement is lowest for diseases of the genitourinary system and mental and behavioural disorders where perfect agreement varies from zero to 4.9% and zero to 12.3% across conditions, respectively. The low agreement between data sources is accompanied by high proportions of individuals with a condition identified only in primary care data (i.e. not in either of the other two sources), with a median of 59.3% (IQR 47.4-75.9) for diseases of the genitourinary system and 66.9% (IQR 42.8-79.2) for mental and behavioural disorders.

Conclusions: Our study highlights the impact of the choice of which data source is used in research on individual LTCs and multimorbidity, and the importance of clearly justifying choices made.