**Background:** Routinely collected healthcare data (e.g., disease registries, primary/secondary care databases, administrative health data) are increasingly linked at the person level for medical research. Linkage can be deterministic (using a unique identifier like the UK NHS number) or probabilistic (using multiple non-unique keys). Linkage error—false matches or missed matches—is unavoidable and can introduce bias of unknown magnitude and direction. Reporting guidelines such as RECORD (2015) and GUILD (2018) emphasize transparency, but adherence in multimorbidity research had not been systematically assessed.
**Methods:** A systematic search of MEDLINE, Embase, and Web of Science was conducted for studies published January 2010–December 2020. Inclusion criteria: multimorbidity research (≥2 specified conditions), use of linked data from ≥2 datasets (≥1 routinely collected), English language, human adults (≥18 years). Titles/abstracts were screened in random order; 20% of full-text papers were reviewed by a second reviewer. Data extraction used a piloted form focusing on data sources, linkage process, quality metrics, and reported issues. A customised 6-domain checklist (Identified as linked routinely collected data, Data source, Linkage variables, Linkage methods, Linkage results, Linkage evaluation) was used to grade reporting quality on a 5-point scale (5=well reported, 1=not reported).
**Key Results:** Of 1,872 initial records, 20 studies were included. Studies came from 10 countries, most commonly the UK (n=8, 40%) and USA (n=4, 20%). Diabetes was the most common condition studied (n=7, 35%). Fourteen studies (70%) received linked data from a trusted third party. Only 3 studies (15%) explicitly reported the linkage method (probabilistic, deterministic, or interactive deterministic). Eight studies (40%) reported the variables used for linkage; only 1 described the quality (completeness/accuracy) of those variables. Only 2 studies reported prelinkage quality checks. Linkage quality metrics were reported by only 3 studies: 2 reported linkage rates (87% and 99.8%), and 1 reported raw linkage figures. Only 1 study checked for bias by comparing patient characteristics of linked vs. non-linked records. None of the 20 studies used statistical methods to adjust for potential linkage error. Five studies reported issues related to the linkage process (e.g., different data source start dates, lack of a unique identifier, data access delays). Eleven studies reported dataset issues: misclassification of disease status (n=7) and missing data (n=8). The overall mean reporting quality score was 2.5 out of 5, indicating only partial reporting. The best-scored domains were 'Identified as linked routinely collected data' and 'Data source'; the worst were 'Linkage variables' (mean 1.5) and 'Linkage evaluation' (median 1). Only 1 study referenced any data linkage reporting guideline (RECORD and GUILD). There was no improvement in reporting quality over time.
**Clinical Implications:** The poor reporting of data linkage processes in multimorbidity research raises serious concerns about unrecognized linkage bias, which can lead to inaccurate estimates of disease clustering, comorbidity burden, and treatment effects. Without transparency about linkage methods, quality, and potential errors, clinicians and policymakers cannot adequately assess the validity of research findings derived from linked datasets. The authors call for increased awareness of linkage bias, better adherence to existing reporting guidelines (RECORD, GUILD), and encouragement from journals and reviewers to include linkage details (at minimum in supplementary materials). The findings are likely generalizable beyond multimorbidity to other fields using linked routinely collected data.