Differing associations with childhood outcomes using behavioural patterns derived from three data reduction techniques

Int J Epidemiol. 2022 Jul 13;dyac142. doi: 10.1093/ije/dyac142. Online ahead of print.

Abstract

Background: Behavioural patterns help to understand the influence of multiple health behaviours on childhood outcomes. Behavioural patterns derived using different data reduction techniques can be non-identical and may differentially associate with childhood outcomes. This study aimed to compare associations of behavioural patterns derived from three methods with three childhood outcomes.

Methods: Data were from the Healthy Active Preschool and Primary Years study when children were 6-8 years old (n = 432). Cluster analysis (CA), latent profile analysis (LPA) and principal component analysis (PCA) were used to derive behavioural patterns from children's diet, physical activity, sedentary behaviour and sleep data. Behavioural data were obtained through parent report and accelerometry. Children's height, weight and waist circumference were measured by trained study staff. Health-related quality of life data were obtained using the Pediatric Quality of Life Inventory and academic performance scores were from a national test. Associations between derived patterns from each method and each of the outcomes were tested using linear regression (adjusted for child age and sex and parent education).

Results: Three patterns were each derived using CA and LPA, and four patterns were derived using PCA. Each method identified a healthy, an unhealthy and a mixed (comprising healthy and unhealthy behaviours together) pattern. Differences in associations were observed between pattern groups from CA and LPA and pattern scores from PCA with the three outcomes.

Conclusions: Discrepancies in associations across pattern derivation methods suggests that the choice of method can influence subsequent associations with outcomes. This has implications for comparison across studies that have employed different methods.

Keywords: Children; academics; cluster analysis; health; health-related quality of life; latent profile analysis; obesity; overweight; principal component analysis.