Marginal structural models: much ado about (almost) nothing.

Marginal structural models were developed to account for a so-called time-dependent confounder and to estimate the presumed effect of 'treatment regime' (treatment over time). We present a set of causal axioms, according to which the problem of time-dependent confounding does not exist, and 'treatment regime' affects nothing. Per our axiomatization, marginal structural models do not introduce a new idea of deconfounding, but simply estimate a weighted average of effects. Whenever a weighted average and the weighting scheme can both be rationalized, the models are acceptable. Whenever a weighted average does not estimate an effect (e.g. important effect modification is ignored), or the weights are senseless - the models should not be fit.