Cross-validation to tune estimate_npmetric_erf()

[m-qin]

What are everyone's thoughts on tuning the bandwidth parameter for estimate_npmetric_erf? Considerations:

1. Leave-one-out cross-validation? Can this be implemented within the estimate_npmetric_erf function?
2. What error function should we use? Page 5 of this paper says that MISE and AMISE (which is the dominating term of MISE) are standard for evaluating kernel density estimators. Interestingly, the bandwidth that optimizes AMISE seems to have a closed form - does that mean that we wouldn't need to use cross-validation at all if we choose AMISE as our error function?
3. Does all this still work if we are using matched or weighted data (i.e., a pseudopopulation), as inputted in the parameter matched_counter?
4. Do these methods still make sense given that the ERF is not a probability density function but simply the function mapping an exposure to an outcome (or perhaps more precisely, the mean outcome over GPS values, if that is correct)?
5. How does estimate_npmetric_erf currently choose the bandwidth from the user-inputted vector bw_seq? From the code, it seems like there is a compute_risk function which is run on the full data for each value of bw_seq?