Propensity Score Matching using MatchIt. How to introduce a caliper to an exact additional specification

Question

I'm new to R, and Im using matchit to match patients by propensity scores. My dataset includes two groups of patients that differ by whether or not they received a drug. As well as matching on propensity score, I also want to exact match on age, gender and family practice, as the decision to treat does vary significantly depending on these characteristics.

My problem is that I would like to exact match by age ± 2 years, as I would like to increase the number of treated patients that can be matched in my dataset, by increasing the flexibility of my exact matches (rather than using an arbitrary cut-off, as with an categorical age variable, for example).

The code below works well

match.1 <-  matchit(exposed ~ imd2010_5 + hyp + mhealth + mm_score + GP_consult + smoke + no_pres3m + lterm_pres + pres_rate, 
data=dataset_pscore2, 
distance=dataset_pscore2$pscores, 
method="nearest", 
exact=c("gender", "pracid", "age_cat"), 
discard="both", 
caliper=0.25, 
ratio=2)

Id just like to change the exact code to something like

exact=c("gender", "pracid", "age" *with a caliper*),

Is this possible?

Answer 1

As far as i can tell, Propensity Score Matching (PSM) usually uses a LOGIT regression to find the estimated propensity scores, as a way to avoid the curse of dimensionality.So, doing exact matching in some covariates in a PSM set-up sounds unlikely.

I suggest running two matching procedures: The first one with exact matching on the covariates which you would like to be exact matched. Then, running the PSM on the new matched-sample on the other covariates.

As an example:

match.1 <-  matchit(exposed ~ gender + pracid + age_cat, 
data=dataset_pscore2, 
method="exact")

matched_sample <- match.data(match.1) # Get Propensity Scores for 'matched_sample'

match.2 <-  matchit(exposed ~ imd2010_5 + hyp + mhealth + mm_score + GP_consult + smoke + no_pres3m + lterm_pres + pres_rate, 
data=matched_sample, 
distance=matched_sample2$pscores, 
method="nearest", 
discard="both", 
caliper=0.25, 
ratio=2)

I also suggest reading the MatchIt package documentation. And this other wonderfull package, called cem.