# stimating the optimal dynamic treatment regime from longitudinal observational data

**Estimating the optimal dynamic treatment regime from longitudinal observational data**

**Orador:** Prof. Liliana Orellana, Facultad de Ciencias Exactas y Naturales - Universidad Buenos Aires**Data:** 24 de Setembro (Sexta-feira), 16:30**Local: **Aula Magna, Faculdade de Medicina da Univ. Porto**Abstract:**

Dynamic treatment regimes, also called adaptive strategies, are individually tailored treatments based on patient covariate history. Optimal dynamic regimes (ODR) are rules that will lead to the highest expected value of some utility function at the end of a time period. Many pressing public health questions are concerned with finding the ODR out of a small set of rules in which the decision maker can only use a subset of the observed covariates. For example, one pressing question in AIDS research is to define the optimal threshold CD4 cell count at which to start prescribing HAART to HIV infected subjects, a rule which only depends on the covariate history through the minimum CD4 count.

In this talk we will present one approach to estimate the ODR when the set of enforceable regimes comprises simple rules based on a subset of

past information and is indexed by a Euclidean vector x. The goal is to estimate the regime g_{x_{opt}} that maximizes the expected counterfactual utility over all enforceable regimes. We will describe how to conduct inference when the expected utility is assumed to follow models that allow the possibility of borrowing information across regimes and across baseline covariates. We consider parametric models on x and on a set of baseline covariates, indexed by a parameter $\beta$. Under these models, the optimal treatment g_{x_{opt}} is a function of $\beta$ and can be estimated using inverse probability weighting.

The proposal is applied to a cohort of HIV positive patients to illustrate estimation of the optimal CD4 count level to start HAART.

Joint work with Andrea Rotnitzky and Jamie Robins.