Abstract: Adherence surveys can be lengthy due to the high number of potential risk factors to be analyzed. As a result, researchers often reduce items into conceptual domains (eg, beliefs, economic factors) to overcome power constraints or focus testing on a specific theme. However, item reduction can also be guided by factor analysis (FA), a process that identifies domains without regard to conceptual frameworks. Although both approaches achieve the same objective, their outputs can be drastically different. It was unclear how the process used to create domains could impact downstream performance of an adherence prediction model. We compared two logistic regression models on the outcome of non-persistence from the same survey data; variables for the models were reduced using a conceptual approach or factor analysis (FA). Both approaches identified three domains from 51 survey items. While domains from the conceptual approach were based on the WHO framework, items contained in FA-guided domains crossed conceptual boundaries. Both models demonstrated good predictive performance with c-statistics of 0.84 (subjective model) and 0.82 (FA model) (p=0.060). The conceptual approach organizes data in a highly relevant structure that aligns with contemporary research and can more readily impact future practice. We found no evidence for a trade-off with respect to model prediction performance. Keywords: factor analysis statistical, medication adherence, assessment of medication adherence, surveys and questionnaires, epidemiologic research design
Umeaghadi et al. (Sun,) studied this question.