Estimation of HTE evolving across single/two covariates

[Tauhmax]

Hi team,

Thank you very much for the nice work! In HTE analysis, typically papers would show how the effect evolve across one/two variables, for the ease to interpret and get insights. For example, given a list of covariates (Age, Gender, Income, ...), we would like to see how a policy affect the health condition differently across these characteristics. And usually we want to check one by one, holding the other covariates constant. I also see a lot papers plot such graph.

Possible solution 1

As I learn from this package, we can use average_treatment_effect function, specifying the subset parameter. Taking age as an example, we can Iteratively feed subgroup with data$Age==18 / data$Age==19 / data$Age==20, ... then collect the output for each subgroup and visualize it. Am I correct on this point?

Possible solution 2

Also, I was wondering if I can use predict to query the effect on each sample, then group these effect based on Age to estimate the CATE for each age. The mean and confidence interval could be derived by aggregating the effect of each sample.

Possible solution 3

Finally, I guess I can use predict(X.test) by manipulating X.test as the characteristics of each group (e.g., Age=18/19/20) and set other column with constant mean/median?

I searched issues and some papers but do not get confirmation on implementation details. Hope to get your comments and suggestions. Thanks!

Best

[erikcs]

Hi @Tauhmax, you are right that there are many approaches you can take to construct visualizations like these. One approach that is popular in practice is to create partial dependence plots by predicting on some chosen X.test, like you hinted at in 3), an example of this is here: https://bookdown.org/stanfordgsbsilab/ml-ci-tutorial/hte-i-binary-treatment.html#partial-dependence

[Tauhmax]

Hi @Tauhmax, you are right that there are many approaches you can take to construct visualizations like these. One approach that is popular in practice is to create partial dependence plots by predicting on some chosen X.test, like you hinted at in 3), an example of this is here: https://bookdown.org/stanfordgsbsilab/ml-ci-tutorial/hte-i-binary-treatment.html#partial-dependence

Thank you @erikcs very much! It is very helpful to have this guidance. Thanks!