2019/01/17/understanding-propensity-score-weighting/

[utterances-bot]

Live Free or Dichotomize - Understanding propensity score weighting

https://livefreeordichotomize.com/2019/01/17/understanding-propensity-score-weighting/

[fallleafsean]

Thank you. It is very clear.
I have a question in "estimating the treatment effect" part. If we want to give a confidence interval for the treatment effect difference of weighting data, e.g. to compare the cure rate (proportion difference) between two treatment(weighted), how to estimate the variance?Is there any reference about this?

[mmarroyo]

Hi Lucy, thank you for the post, really useful and easy to understand.

I have a question regarding propensity score methods. I have also heard about Inverse probability score weighthing (IPTW), how do this calculation fit into this explanation? Is IPTW a type of propensity score method?

KR,
Marta

[LucyMcGowan]

Hi Marta,

These are all examples of probability weighting in this post; most often when papers refer to IPTW they are talking about ATE weights (so weighting by the inverse of the propensity to receive the treatment you received)

[tanguy0807]

Thank you for the post Lucy, the graphs help understand the different kind of weighting strategies.

However, I was wondering why propensity scores were even being used in the first place in your example. I thought propensity scores were useful when one confounders should be controlled for (and, if I am correct, only for confounders should be included in the propensity model) but in the case of your synthetic dataset I fail to see any. I understand that variables x_1 and x_2 cause treatment, and only the latter causes outcome (so x_1 and x_2 aren't confounders). So couldn't the analysis be performed without using x_1 and x_2 (at least for the ATE)?

[LucyMcGowan]

Oh goodness @tanguy0807, thank you! I definitely intended for these to be confounders, I've updated the simulation to match that.