Measuring implementation fidelity in a cluster-randomized pragmatic trial: development and use of a quantitative multi-component approach

Background: In pragmatic trials, on-site partners, rather than researchers, lead intervention delivery, which may result in implementation variation. There is a need to quantitatively measure this variation. Applying the Framework for Implementation Fidelity (FIF), we develop an approach for measuring variability in site-level implementation fidelity. This approach is then applied to measure site-level fidelity in a cluster-randomized pragmatic trial of Music & MemorySM (M&M), a personalized music intervention targeting agitated behaviors in residents living with dementia, in US nursing homes (NHs). Methods: Intervention NHs (N = 27) implemented M&M using a standardized manual, utilizing provided staff trainings and iPods for participating residents. Quantitative implementation data, including iPod metadata (i.e., song title, duration, number of plays), were collected during baseline, 4-month, and 8-month site visits. Three researchers developed four FIF adherence dimension scores. For Details of Content, we independently reviewed the implementation manual and reached consensus on six core M&M components. Coverage was the total number of residents exposed to the music at each NH. Frequency was the percent of participating residents in each NH exposed to M&M at least weekly. Duration was the median minutes of music received per resident day exposed. Data elements were scaled and summed to generate dimension-level NH scores, which were then summed to create a Composite adherence score. NHs were grouped by tercile (low-, medium-, high-fidelity). Results: The 27 NHs differed in size, resident composition, and publicly reported quality rating. The Composite score demonstrated significant variation across NHs, ranging from 4.0 to 12.0 [8.0, standard deviation (SD) 2.1]. Scaled dimension scores were significantly correlated with the Composite score. However, dimension scores were not highly correlated with each other; for example, the correlation of the Details of Content score with Coverage was tau(b) = 0.11 (p = 0.59) and with Duration was tau(b) = - 0.05 (p = 0.78). The Composite score correlated with CMS quality star rating and presence of an Alzheimer's unit, suggesting face validity. Conclusions: Guided by the FIF, we developed and used an approach to quantitatively measure overall site-level fidelity in a multi-site pragmatic trial. Future pragmatic trials, particularly in the long-term care environment, may benefit from this approach.