threshold in propensity score matching

[kosukeimai]

threshold in propensity score matching

[ngreifer]

There are no thresholds involved, so I'm not sure what you mean. The matched data contains all the units that have been matched. Nearest neighbor matching involves find the nearest control unit to each treated unit and discarding the rest.

You might be thinking of a caliper, which is a limit on how far apart units can be to be able to be matched with each other. You can specify that with the caliper argument, but that is not a necessary part of matching. Please read Austin (2011) and Ho, Imai, King and Stuart (2007) to understand how to conduct a matching analysis and what MatchIt does.

[yceny]

Thanks for your response. I read the two papers. They are very useful. Yes, caliper is what I mean here. One more clarification question. In matchit function, if we specify, say caliper = 0.2, does this mean 0.2 of the standard deviation of the logit of the propensity score?

[ngreifer]

Setting caliper = .2 will by default use a caliper of .2 standard deviations of the distance measure. If the distance measure is the propensity score, then it will be .2 standard deviations of the propensity score. To make it the logit of the propensity score, use link = "linear.logit" in the call to matchit(). This will make the logit of the propensity score the distance used for matching and the variable to which the caliper is applied.