The impact of physical activity on psychological distress in young people: An analysis of longitudinal data from the Millennium Cohort Study using causal machine learning

Physical activity is often associated with positive psychological wellbeing in adolescents, and there is evidence from randomised trials to support this. However, the effect sizes reported from intervention studies vary considerably. Moreover, physical activity is a 'complex intervention' and findings from trials may not generalise to naturalistic settings. Estimating causal effects from observational data is challenging, though provides an opportunity to add to the evidence base in this area. We used newer machine learning-based causal inference methods to estimate (i) the causal effect of physical activity on psychological distress in young people (ii) variations in effects across subgroups ('heterogeneous treatment effects').

Data from the Millennium Cohort Study were available for 9123 adolescents. Self-reported days of moderate/vigorous activity was recorded at ages 14 and 17. Psychological distress was measured at age 17 using the Kessler-6 scale. Directed acyclic graphs were co-produced with a youth advisory group and individuals with lived experience to identify confounders. Average and heterogeneous treatment effects of physical activity at age 14 on psychological distress at age 17, as well as concurrently at age 17, were estimated using causal machine learning methods (targeted learning, causal forests).

We observed no overall impact of physical activity at age 14 on self-reported distress scores at age 17 (0.11 points, -0.09 to 0.31, p = .28), although there was a modest impact of physical activity at 17 on concurrent distress (-0.39, -0.61 to -0.18, p < .001). Meeting government guidelines on physical activity levels benefited boys more than girls, but had relatively less positive impact on young people with special educational needs.

Whilst no overall impact of physical activity levels on youth distress was observed, boys may particularly benefit from targeted interventions in this regard. Causal machine learning is a promising approach to rapidly generating evidence from observational data.