2024-08-06
\[P(y|\text{do}(x)) = \sum_z P(y|x,z)P(z)\]
assume the following structural causal model (\(z\) is confounder, \(u\) is exogenous noise): \[f_y(t,z,u) = \beta_t t + \beta_z z + \beta_u u\]
then: \[\begin{align} \text{ATE} &= E[Y|\text{do}(t=1)] - E[Y|\text{do}(t=0)] \\ &\class{fragment}{= E_{z,u}[\beta_t * 1+ \beta_z z + \beta_u u] - E_{z,u}[\beta_t * 0 + \beta_z z + \beta_u u]} \\ &\class{fragment}{= \beta_t + E_{z,u}[\beta_z z + \beta_u u] - E_{z,u}[\beta_z z + \beta_u u]} \\ &\class{fragment}{= \beta_t} \end{align}\]
i.e. the ATE collapses to the the regression parameter \(\beta_t\) in a linear regression model of \(y\) on \(t,z\)
The number of possible DAGs grows super-exponentially in the number of nodes
n_nodes | n_dags | time at 1 sec / DAG |
---|---|---|
1 | 1 | |
2 | 3 | |
3 | 25 | |
4 | 543 | |
5 | 29281 | > an hour |
6 | 3781503 | > a day |
7 | 1138779265 | > a year |
8 | 783702329343 | |
9 | 1213442454842881 | > human species |
10 | 4175098976430598143 | > age of universe |
Wouter van Amsterdam — WvanAmsterdam — vanamsterdam.github.io