Biostatistics

In the past five years the team worked on two main topics: multivariate models for time-dependent data and model-based estimation of public health indicators. Our main domain of research focus on the development of multivariate dynamic models for the analysis of censored time-to-events and/or repeated measures of longitudinal data accounting for complex observation schemes.

These works are motivated by the study of the natural history of chronic diseases such as the Alzheimer disease or the Multi-System Atrophy, the investigation of the impact of time-dependent exposures, or the validation of surrogate markers for clinical trials in cancer research.

Parametric and semiparametric estimation procedures for frailty models for correlated time-to-events, clustered data and/or recurrent events as well as joint models for event times and longitudinal markers were implemented in the R-package Frailtypack. Another field of research regards the extension of mixed models using latent classes and/or latent processes for the analysis of multiple longitudinal outcomes with non-standard distributions in heterogeneous populations.

We proposed the R-package LCMM that enables the estimation of latent class mixed models, joint latent class mixed models and mixed models for curvilinear univariate or multivariate longitudinal outcomes. These models were motivated by the analysis of cognitive decline in cohort studies.

They account for the population heterogeneity and the issues raised by the metrologic properties of measurement tools of cognition and autonomy (high correlation between markers measuring one or several underlying processes, ordinal data, non-standard asymmetric distributions with floor and/or ceiling effects and unequal sensitivity to changes).

We also developed methods for the estimation of Illness-Death model accounting for interval-censoring (Package SmoothHazard). Tools for computing individual prediction and evaluating predictive abilities of these models are also implemented.

Relying on multi-state methodology, we proposed several approaches to forecast the future burden of neurologic diseases and assess the expected impact of intervention scenarios targeting their modifiable risk factors.

Depending on the complexity of the investigated scenarios, the indicators for the future burden of the disease are computed analytically or using micro-simulations.

Our current projects particularly focus on causal questions and big-data issues in the framework of dynamic models. On one hand, causal questions are related to our research about the mechanism underlying pathological processes in chronic diseases, the role of some long-term exposure and the impact of social inequalities in health.

We will investigate the causal interpretation of the multivariate models we develop and we will propose new methods for studying causality for censored time-to-events and/or time-dependent risk factors. On the other hand, as technological progress allow collecting a very large amount of data (genetics, biology, imaging, IoT data), we investigate new approaches that tackle high-dimensionality issues with respect to the number of time-dependent predictors, markers and outcomes.