Matching and data analysis
ABCG2 c.421C>A variant carriers and wt controls were
matched in combined exact and optimal full matching with Mahalanobis as
a distance measure using package MatchIT [21] in R [19].
The procedure allows “one-to-many” variant to control matching (andvice-versa ) and attains (exact) or approximates (Mahalanobis)
balance achieved by fully blocked randomization (in respect to measured
confounders) (see ESM – Supplemental Methods B, for details)
[21-23]. Since in substantially different ranges (70-341 vs.
3.4-37.4 µg/L), to be used in matching CsA and tacrolimus troughs were
rescaled [ln(tacrolimus) troughs rescaled to ln(CsA troughs) range by
linear transformation]. Inadequately matched covariates (standardized
mean difference, d ≥0.1) were adjusted for in data analysis. The variant
allele effect on (ln-transformed) pharmacokinetic outcomes was estimated
in raw and matched/adjusted data in frequentist (maximum likelihood with
Gauss-Hermite approximation for raw data; cluster robust variance
estimator for matched data) and Bayesian (4 chains, 4000 iterations,
8000 samples of the posterior, highest posterior density [HPD]
credible intervals) general linear models, and was expressed as
geometric means ratio (GMR). In the latter, we defined a moderately
informed skeptical prior for the effect of interest consistent with thea priori hypothesis of no effect: centered at 0 for ln(GMR) with
a standard deviation of 0.355. It assigns with 95% probability to a GMR
between 0.5 and 2.0, and 48% probability to a GMR within the
“conventional” limits of equivalence (0.80 to 1.25). We used SAS 9.4
for Windows (SAS Inc., Cary, NC) to fit frequentist models and R packagenstanarm [24] to fit Bayesian models. We used CubeX [25]
to evaluate linkage disequilibrium (LD).