Here is an idea. What if we use linear regression?
Let's say you fit the following linear model to your data.
$y_i = \beta_0 + \beta_1 t_i + \beta_2 X_i + e_i$
Should not it be mentioned here or maybe in some other place in this chapter that features $X$ are expected to be good controls, as defined in Good Controls section of the book, since sensitivity defined as $\frac{\delta y_i}{\delta t_i} = \beta_1$ would have casual meaning when linear regression is capable to find ATE, i.e. it is constructed according to the casual model? Or if generally speaking $X$ is just a set of observable variables, there is no guarantee the linear regression will find an ATE, and $\beta_1$ is just some regression coefficient, which would not answer the question of what effect in $y$ we see had we changed $t$ by $\epsilon$.
There is an issue on chapter 18 - Heterogeneous Treatment Effects and Personalization, in the following paragraph Predicting Sensitivity
Should not it be mentioned here or maybe in some other place in this chapter that features $X$ are expected to be good controls, as defined in Good Controls section of the book, since sensitivity defined as $\frac{\delta y_i}{\delta t_i} = \beta_1$ would have casual meaning when linear regression is capable to find ATE, i.e. it is constructed according to the casual model? Or if generally speaking $X$ is just a set of observable variables, there is no guarantee the linear regression will find an ATE, and $\beta_1$ is just some regression coefficient, which would not answer the question of what effect in $y$ we see had we changed $t$ by $\epsilon$.