A comparison of children’s diet and movement behaviour patterns derived from three unsupervised multivariate methods

BACKGROUND: Behavioural patterns are typically derived using unsupervised multivariate methods such as principal component analysis (PCA), latent profile analysis (LPA) and cluster analysis (CA). Comparability and congruence between the patterns derived from these methods has not been previously investigated, thus it’s unclear whether patterns from studies using different methods are directly comparable. This study aimed to compare behavioural patterns derived across diet, physical activity, sedentary behaviour and sleep domains, using PCA, LPA and CA in a single dataset. METHODS: Parent-report and accelerometry data from the second wave (2011/12; child age 6-8y, n = 432) of the HAPPY cohort study (Melbourne, Australia) were used to derive behavioural patterns using PCA, LPA and CA. Standardized variables assessing diet (intake of fruit, vegetable, sweet, and savoury discretionary items), physical activity (moderate- to vigorous-intensity physical activity [MVPA] from accelerometry, organised sport duration and outdoor playtime from parent report), sedentary behaviour (sedentary time from accelerometry, screen time, videogames and quiet playtime from parent report) and sleep (daily sleep duration) were included in the analyses. For each method, commonly used criteria for pattern retention were applied. RESULTS: PCA produced four patterns whereas LPA and CA each generated three patterns. Despite the number and characterisation of the behavioural patterns derived being non-identical, each method identified a healthy, unhealthy and a mixed pattern. Three common underlying themes emerged across the methods for each type of pattern: (i) High fruit and vegetable intake and high outdoor play (“healthy”); (ii) poor diet (either low fruit and vegetable intake or high discretionary food intake) and high sedentary behaviour (“unhealthy”); and (iii) high MVPA, poor diet (as defined above) and low sedentary time (“mixed”). CONCLUSION: Within this sample, despite differences in the number of patterns derived by each method, a good degree of concordance across pattern characteristics was seen between the methods. Differences between patterns could be attributable to the underpinning statistical technique of each method. Therefore, acknowledging the differences between the methods and ensuring thorough documentation of the pattern derivation analyses is essential to inform comparison of patterns derived through a range of approaches across studies.