Using MatchIt to match on percentiles of propensity scores while exact matching

Question

I'm using MatchIt to perform time-varying propensity score matching. I estimate propensity scores and do nearest neighbour matching on those, as well as exact matching on a few variables (some which are used in the estimation of propensity scores, some which are not), as follows:

  matches <- matchit(
    
    ## Estimate propensity scores and perform nearest neighbour matching on propensity scores
    y ~ x1
      + x2
      + x3
      + x4
      + x5
    
    , method = "nearest" # matching method: nearest neighbour matching (on propensity score) 
    , distance = "glm" # method for estimating the propensity score: 'glm' = logit
    
    # Also perform exact matching on additional variables
    , exact = ~ x3 
              + x4 
              + x6
  
    , data = df
    , s.weights = ~ sampling_weights 
  )

Balance is good across all variables, but it's not so good on the propensity scores.

I think that matching on percentiles of propensity scores would solve this problem. My understanding is that this could be achieved by changing the 'method' argument to:

method = "subclass", subclass = 100

However, I don't think it's possible to use method = subclass while exact matching on other variables.

Can anyone say if it's possible to match on percentiles of propensity scores, while exact matching on other covariates using MatchIt?

Edited for clarity

There seems to be a slight misconception of terms, "exact" means to find an observation that is _exactly_ identical on selected covariates. — jay.sf, Sep 23 '22 at 14:41

score 0 · Accepted Answer · answered Sep 30 '22 at 06:38

Balance is good across all variables, but it's not so good on the propensity scores.

Balance doesn't need to be good on the propensity scores. In fact, Stuart et al. (2013) found that balance on the propensity score is totally uncorrelated with bias. The purpose of matching is to achieve balance on the covariates; the propensity score is just an instrument to achieve that end. This is the propensity score tautology described in Ho et al. (2007). It sounds like your nearest neighbor match would be sufficient if balance was achieved, though it also sounds like your results might improve with a more sophisticated matching method, like genetic matching. Remember that many matching methods don't involve a propensity score at all.

You can also try full matching, which is very similar to subclassification with many subclasses and does allow exact matching.

If you can tell me what you think subclassification with exact matching constraints is supposed to look like, I can tell you how to achieve it. But subclassification works differently from other matching methods and it's not immediately clear how to combine it with subclassification.

Ho, D. E., Imai, K., King, G., & Stuart, E. A. (2007). Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference. Political Analysis, 15(3), 199–236. https://doi.org/10.1093/pan/mpl013

Stuart, E. A., Lee, B. K., & Leacy, F. P. (2013). Prognostic score-based balance measures can be a useful diagnostic for propensity score methods in comparative effectiveness research. Journal of Clinical Epidemiology, 66(8), S84. https://doi.org/10.1016/j.jclinepi.2013.01.013

Thanks Noah, this is very helpful. The fact that balance on the propensity scores doesn't matter makes a lot of sense now that you say it. Thanks a lot for those references too. I've had a look at genetic matching but the balance achieved using nn is good, so I think sticking with nn makes sense. Thanks again. Love MatchIt :) — shiny_boson, Oct 07 '22 at 13:08

Using MatchIt to match on percentiles of propensity scores while exact matching

1 Answers1