We have a longitudinal experiment, with interventions $\bar{A}=\{A_1,A_2,\ldots,A_K\}$ and outcomes $\bar{Y}=\{Y_1,Y_2,\ldots,Y_K\}$. Using some sequential conditional exchangeability assumptions, it should be possible to derive g-computation formula to estimate the causal effect of a treatment strategy $\bar{a}$.

Conditional exchangeability commonly assumed for a longitudinal trial are

$Y_{k'}^{\bar{a}}\ \bot A_k\ |\ {\bar{A}}_{k-1}={\bar{a}}_{k-1},{\bar{Y}}_k$ for $k=0,1,\ldots,k'$

I.e., we need to condition on treatments and observations occurring prior to $A_k$.

The resulting g-computation is

$$ p({\bar{Y}}^{\bar{a}})\ =\ \prod_{k=1}^{K}{p\left(Y_k\ \middle|\ \bar{Y}_{k-1},\bar{A}_{k-1}\right)} $$

How can this be derived?

I’m able to derive the g-computation, but only if I assume

$Y_{k\prime}^{\bar{a}}\bot A_k|{\bar{A}}_{k-1}={\bar{a}}_{k-1},{\bar{Y}}_{k'-1}$ for $k=0,1,\ldots,k'$

I.e., condition on treatments occurring prior to $A_k$ but observations prior to $Y_{k'}^{\bar{a}}$. Thus, condition also on observations that occur after the treatment $A_k$ but before the observation $Y_{k'}$. Details are written up at https://doi.org/10.6084/m9.figshare.20406687. It remains that I cannot derive the commonly used g-computation from the commonly made sequential conditional exchangeability assumptions, and I would appreciate help.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.