Supplemental Methods B – Matching
In respect to variables that needed to be controlled for (in addition to
inclusion-exclusion criteria), the present setting was extremely
complex. For example, based on categorical covariates depicted in Figure
S1, 768 strata of variable combinations could be formed. We used exact
matching combined with optimal full matching based on Mahalanobis
distance with age, BMI, eCrCl and CNI concentrations as (further)
continuous matching variables. Exact matching achieves balance that
corresponds to a fully blocked randomization. Optimal full matching also
allows one-to-many matching, and when Mahalanobis is a distance measure,
it approximates fully blocked randomization. The process first completes
exact matching on specified covariates (forms subclasses of “treated”
and “controls” exactly matched on a set of covariates) and then uses
Mahalanobis distance (calculated using the entire data set) to further
minimize within-subclass distances regarding covariates not included in
exact matching [1-3]. Choice of (categorical) covariates for exact
matching was guided by the intention to retain as many as possible
“treated” subjects, but also on their practical importance: hence,
exact matching was in respect to the type of CNI (CsA or tacrolimus),UGT2B7 and UGT1A9 SNPs.