search

ngreifer / WeightIt

WeightIt: an R package for propensity score weighting
https://ngreifer.github.io/WeightIt/
102 stars 12 forks source link

Sampling weights: to trim or not to trim #55

Closed lorenzoFabbri closed 8 months ago

lorenzoFabbri commented 8 months ago

This is another mostly conceptual question, and not a bug/issue with the package.

I am doing sensitivity analysis on a small subset of my sample. I thus decided to estimate the weights to reduce the effects of any selection bias:

selection ~ covariates

where selction is a indicator variable which is 1 when the subject was indeed selected into the subset, 0 otherwise. I then pass the weights from the call to WeightIt::weightit to s.weights when estimating the actual balancing weights.

I was wondering whether it is better to eventually trim the weights for selection bias, or if it is better to pass them as is to the second step and then trim these final balancing weights. I know I can check the summary and see whether there are extreme weights, but I was looking for a more conceptual understanding. Thank you.

ngreifer commented 8 months ago

I assume by trimming you mean setting extreme weights to a less extreme value. If so, that is just a modeling decision, no different from choosing probit vs logit regression. So you can trim at any point in the analysis, just the same as you would by choosing a different model. Do whatever it takes to achieve balance and a high ESS. There are no rules otherwise.