Seminar in applied mathematics and statistics
SPEAKER: Helene Charlotte Rytgaard (Biostatistics, University of Copenhagen).
TITLE: Targeted causal inference from time-varying observational data.
ABSTRACT: Observational longitudinal data such as provided by the Danish nation-wide registries are routinely used to estimate treatment-outcome associations. Standard methods such as Cox regression with time-varying exposure are limited by difficulties such as the interpretation of the hazard ratio as the measure of treatment effect. In many applications, it is thus of interest to formulate and estimate the effect of dynamic treatment regimes on the absolute risk scale.
Targeted minimum loss-based estimation (TMLE) provides a general template for constructing regular and asymptotically linear substitution estimators for smooth low-dimensional parameters in semiparametric models. Procedually, a TMLE is constructed in two steps that combine flexible ensemble learning and semiparametric efficiency theory. While most research on TMLE methods has focused on data given on a discrete time-scale, observational data rarely consists of measurements made on a regular time-grid.
In our work, we take a continuous-time approach. We utilize a counting process framework, and define our target of estimation as the intervention-specific mean outcome at the end of follow-up in a semiparametric model. We derive the efficient influence function for the statistical estimation problem. This provides the basis for construction of an efficient estimator for which further double robustness and statistical inference can be established. In the talk, I will give an insight into our ideas for doing TMLE in the continuous-time setting.
Upcoming events (after November 13):
Wednesday, November 20 at 15.15: Sacha Desmettre
Wednesday, November 27 at 15.15: Jesper Møller
Wednesday, November 27 at 16.15: Carl Graham
Wednesday, February 19 at 15.15: Budhi Surya