Forum de statistique : conférence "Meta-analysis of heterogeneous predictive models based on Monte Carlo simulations and the concept of a meta-model"

- Forum de statistiques
M. Aziz Chaouch

Division Biostatistique et méthodes quantitatives (dBMQ)
IUMSP, Lausanne

Abstract: Predictive models are being increasingly used in clinical practice e.g. to predict future drug plasma concentrations in subpopulations of patients as a function of the dosage regimen. When different studies developed predictive models for the same outcome, one would like to combine evidence from these models to optimize the quality of predictions. One problem is that the different studies may have used different predictive models with e.g. different sets of predictors. We shall present a simulation approach to allow a meta-analysis of such heterogeneous predictive models. The method postulates a hypothetic model (called the meta-model) which supposedly approximates the data generating mechanism common to all studies. By generating pseudo-observations from the meta-model and refitting the possibly misspecified models to these pseudo-data, it is possible to estimate meta-model parameters using an iterative procedure. Statistical properties of the method will be presented in the context of linear and linear mixed models.

This is a joint work with Prof.Thierry Buclin and Prof. Valentin Rousson

Date et heure
Mardi, 15 Mai, 2018 - 11:00
Salle Delachaux (étage 01), Institut universitaire de médecine sociale et préventive, Bâtiment Biopôle 2, route de la Corniche 10, 1010 Lausanne
Série de conférences

Retrouver sur notre site


Institut universitaire de médecine sociale et préventive
Route de la Corniche 10, 1010 Lausanne - Switzerland
+41 21 314 72 72 |

Go to top