Targeted Learning

Abstract

Talk in Spanish. This seminar presents an introduction to targeted learning (TL). TL is a statistical tool of recent development that can be applied to the analysis of data corresponding to continuous or categorical variables, coming from both observational and experimental studies. Theoretical aspects will be illustrated for the estimation of a targeted parameter. In this case, a prediction index of mortality risk to answer the objective question about the effect of physical activity on the mortality rate of older adults. TL consists of two steps. First, an efficient initial estimate is obtained through a procedure called super learning (SL). To obtain the initial estimate, TL simultaneously uses a wide variety of models and algorithms. The methods incorporated by SL come from both classical statistics and machine learning, which is more common in computer science. SL uses a loss function to choose the most efficient linear combination of estimates using different models consistent with the structure of the data. Through the use of cross validation, SL prevents overfitting. The second step of TL is a correction of bias in the initial estimate. This is done through a procedure called maximum likelihood estimation. TL provides estimates with an optimal balance between bias and variance in the estimation of the objective parameter. It is possible to make inferences On these estimates. Despite being a tool developed in the last decade, TL has a solid theoretical and empirical basis that demonstrates the equality and, frequently, superiority of TL when compared with other analytical tools. TL is implemented in the SuperLearner and tmle R packages.