In 1986, Robins introduced a new framework for drawing causal inference from observational data. In this and other articles published around the same time, Robins showed that in non-experimental data, exposure is almost always time-dependent, and that standard methods such as regression are therefore almost always biased. This framework is mathematically very closely related to
Judea Pearl's graphical framework Non-Parametric Structural Equations Models, which Pearl developed independently shortly thereafter. Pearl's graphical models are a more restricted version of this theory. In his original paper on causal inference, Robins described two new methods for controlling for confounding bias, which can be applied in the generalized setting of time-dependent exposures: The G-formula and G-Estimation of Structural Nested Models. Later, he introduced a third class of models,
Marginal Structural Models, in which the parameters are estimated using inverse probability of treatment weights. He has also contributed significantly to the theory of dynamic treatment regimes, which are of high significance in
comparative effectiveness research and personalized medicine. Together with
Andrea Rotnitzky and other colleagues, in 1994 he introduced doubly robust estimators (derived from the influence functions) for statistical parameters in causal inference and missing data problems. The theory for doubly robust estimators has been highly influential in the field of
Causal_inference and has influenced practice in computer science, biostatistics, epidemiology, machine learning, social sciences, and statistics. In 2008, he also developed the theory of higher-order influence functions for statistical functional estimation with collaborators including Lingling Li, Eric Tchetgen Tchetgen, and
Aad van der Vaart. == Selected publications ==