A call for caution when using network methods to study multimorbidity: an illustration using data from the Canadian Longitudinal Study on Aging.

ABSTRACT

OBJECTIVES: To examine the impact of two key choices when conducting a network analysis (clustering methods and measure of association) on the number and type of multimorbidity clusters. STUDY DESIGN AND SETTING: Using cross-sectional self-reported data on 24 diseases from 30,097 community-living adults aged 45-85 from the Canadian Longitudinal Study on Aging, we conducted network analyses using 5 clustering methods and 11 association measures commonly used in multimorbidity studies. We compared the similarity among clusters using the adjusted Rand index (ARI); an ARI of 0 is equivalent to the diseases being randomly assigned to clusters, and 1 indicates perfect agreement. We compared the network analysis results to disease clusters independently identified by two clinicians. RESULTS: Results differed greatly across combinations of association measures and cluster algorithms. The number of clusters identified ranged from 1 to 24, with a low similarity of conditions within clusters. Compared to clinician-derived clusters, ARIs ranged from -0.02 to 0.24, indicating little similarity. CONCLUSION: These analyses demonstrate the need for a systematic evaluation of the performance of network analysis methods on binary clustered data like diseases. Moreover, in individual older adults, diseases may not cluster predictably, highlighting the need for a personalized approach to their care.