For making inference on the causal effects on response of time-dependent treatments from observational data where time-dependent confounders may exist, Robins has proposed two classes of models. Marginal structural models (MSMs, discussed in the talk on 2/11/00) are one such approach, but these have some limitations. The other approach is based on a class of probability models Robins calls Structural Nested Models (SNMs). These models offer an advantage over other approaches to making causal inference in such situations, e.g. use of the G-computation algorithm to estimate causal effects, in that they avoid the so-called "null paradox" (discussed in the last talk on 3/30/00) and are more flexible than MSMs in terms of the inferences that may be drawn on causal effects.This talk will provide an introduction to SNMs and how causal inference in these complex situations may be based on them.
Return to Biostatistics Working Group