Abstract
BACKGROUND: Patient self-care and caregiver contribution to self-care in chronic illnesses should be considered together as a dyadic phenomenon called "dyadic illness management." The possibility of classifying dyadic engagement in chronic conditions care may uncover behavioral patterns useful for improving disease management. Mixed-effects models have been used to obtain dyadic scores, which serve as inputs for latent class analysis. However, the advantages of this approach over simpler synthetic dyadic measures remain unclear. OBJECTIVE: The aim of this study was to compare two methods for obtaining dyadic scores to serve as inputs for identifying distinct patterns of dyadic illness management for type 2 diabetes (T2D) through latent class analysis. METHODS: This work uses data from a cross-sectional study on 251 patients with T2D and their informal caregivers. Patient self-care and caregiver contribution to self-care were measured by the Self-Care of Diabetes Inventory and the Caregiver Contribution to Self-Care of Diabetes Inventory, respectively. To assess dyadic illness management, we first adopted the incongruence model, a mixed-effects model with a specific codification that enables the estimation of both the average and incongruence in the outcome within each dyad, through random intercepts and random slopes. These estimated coefficients were then used to perform a latent class analysis that was able to identify patterns of dyadic management. As an alternative approach, we calculated the dyadic average and incongruence using the raw means and the difference between patient and caregiver scores. These values were then used as inputs for latent class analysis. RESULTS: The latent class analysis clustered the same dyads into the same classes across both approaches, with identical fit indices. The three-class model showed the best performance in terms of both fit and characterization of the dyads. DISCUSSION: Mixed-effects models account for interdependence within the dyad and measurement error, returning predicted measures that are shrunk toward the overall mean. However, this approach yielded the same clusters as the simpler method based on observed measures.