Life course epidemiology: Individual causality from disruptive, heterogeneous & multidimensional temporal data
Description: Life course epidemiology (LCE) investigates the interplay of many life-time risk factors, time-dependent events, and health and disease outcomes in individuals. It is an extension of more traditional epidemiology to be able to move towards a systemic and causal understanding of population health and disease. LCE requires longitudinal data on a large number of individuals over a longer period of life, namely multiple temporal risk factors, exposures, and health variables in combination with several later life health or disease outcomes. Modeling all of this data simultaneously has been a long-standing struggle within the field of statistics and epidemiology due to data complexities: the high dimensionality of data and variables, complex correlation structures among variables within and over time, and subject-specific heterogeneities and time-varying events and interventions. The proposed research is a step forward towards solving this struggle through the development of three innovative and sophisticated joint models, sequentially increasing in mathematical complexity. These models integrate the commonly used trajectory or pathway analysis in LCE and they have the benefit of an individual causal interpretation that is in line with the well-known deterministic causality theory of the sufficient cause model. Our models may be applicable to many longitudinal data sets (e.g. cohort studies, administrative data, hospital records of chronic patients, long-term data of value-based healthcare programs), but we will illustrate the relevance of our approach on the longitudinal and life course data of the Framingham Heart Study.
The first joint model will implement piece-wise linear trajectories for a single time-varying risk factor, to incorporate the influence of a time-dependent treatment, with subject-specific heteroskedastic residuals. This latent variable model will then be connected to a survival time model for a single later-life event time. The second joint model extents the first joint model with a second longitudinal risk factor, and the final joint model will add a competing risk to the survival model of the second joint model. Thus our most sophisticated proposed joint model addresses bivariate risk factor time-profiles with random heteroskedastic residuals, time-dependent profile disruptions, and integrated competing risk survival models for later life events. The most sophisticated current joint models do not address any of these complexities and our research will therefore directly and foremost add to the statistical science.
Our contribution also makes it possible to provide a systemic and causal view created from highly disruptive, heterogeneous, and multidimensional temporal data. Indeed, our joint models are related to a directed acyclic graph (DAG) that describes the pathways and life course of health and disease at the population level. It also extents marginal structural models for estimation of just the time-varying treatment effects. Thus our research will also have implications for the field of epidemiology. Finally, we will construct mathematical conditions (i.e. the distributional or DAG restrictions) for our joint models to have a causal interpretation at the individual level. Literature has shown causal interpretation of time-varying treatment effects on longitudinal health outcomes with mixed models when their latent variables are not part of the probability of the time-dependent treatment. These conditions also allow for the estimation of an individual causal effect (through the best linear unbiased predictions), although this has not yet been recognized explicitly. We will extend these known conditions to our sophisticated joint modeling with multiple risk factors and competing events and time-stationary confounders. This may help revive and operationalize the research area on individual causality, which was the basic feature or aspect of a deterministic approach towards individual causality (i.e. the sufficient cause model).
Funding type: grant
Funder scheme: Open Competitie ENW - M Doorlopend OC ENW KLEIN pakket 04
Start date: 2020-08-01
Investigator: Edwin Heuvel
Institution: Technische Universiteit Eindhoven - Eindhoven University of Technology
Funded outputs count: 0
Funder award ID: OCENW.KLEIN.185
OpenAlex ID: G9000095809