A longitudinal causal graph analysis investigating modifiable risk factors and obesity in a European cohort of children and adolescents

Abstract

A Cohort Causal Graph (CCG) over the life-course from childhood to adolescence is estimated to identify potential causes of obesity and to determine promising targets for prevention strategies. We adapt a popular causal discovery algorithm to deal with missing values by multiple imputation and with temporal cohort structure. To estimate possible causal effects of modifiable risk factors at baseline on obesity six years later, we used the “Intervention-calculus when the Directed Acyclic Graph is Absent” and double machine learning with confounder adjustment based on the obtained CCG.Causal relations among 51 variables were analysed including obesity, early life factors, lifestyle and cultural background of 5,112 children from the European IDEFICS/I.Family cohort across three waves (2007-2014). The resulting CCG shows some but not many and only indirect possible pathways from earlier modifiable risk factors such as audio-visual media consumption (AVM) to later obesity. The estimated causal effects suggested that promising interventions would encourage longer sleep or reduce AVM during childhood, both slightly decreasing expected body mass index six years later. But overall, no or only weak causal effects could be found for hypothetical interventions on individual behaviors in early childhood on later obesity.

Inferring Causality Is Preference-Sensitive: We Need a Book of Who as Well as Why

“…Setting the boundaries of a DAG -deciding what nodes/variables are included in the diagram -is an irredeemably human exercise. Impressive DAG-based machine learning exercises may contain large numbers of variables, for example, 29 in Ramsay [5] and 51 in Foraita's innovatory longitudinal Cohort DAG [6]. But the issues focused on here are not about the nodes included or not, but who made these decisions.…”

Wearables and smartphones for modifiable risk factors in metabolic health: a scoping review protocol

“…These include lifestyle determinants (i.e., nutrition, physical activity, sleep, stress, and substance abuse) and physiological markers (i.e., blood sugar, triglycerides, and HDL cholesterol) [6–10]. Following a complex systems perspective, modifiable factors interact dynamically [11] and together shape disease outcomes over time, with up to 70% of cardiovascular disease cases and mortality attributed to modifiable risk factors [12] [9].…”

“…These include lifestyle determinants (i.e., nutrition, physical activity, sleep, stress, and substance abuse) and physiological markers (i.e., blood sugar, triglycerides, and HDL cholesterol) [6][7][8][9][10]. Following a complex systems perspective, modifiable factors interact dynamically [11] and together shape disease outcomes over time, with up to 70% of cardiovascular disease cases and mortality attributed to modifiable risk factors [12] [9]. In parallel, wearables (i.e., devices worn on the wrist, finger, arm, and chest) and smartphones are increasingly used to track modifiable risk factors in daily life with improved precision and accuracy [13].…”

Multiple imputation and test‐wise deletion for causal discovery with incomplete cohort data

“…Figure 1 visualizes the missingness pattern in a subsample of the IDEFICS data containing 657 children from Germany. The choice of variables roughly follows Foraita et al, 26 who performed causal discovery on a larger subsample of the IDEFICS and I.Family data including more children and time points. See Table 1 for details on the variables used in the present paper.…”