Small-sample and selection bias effects in multivariate calibration, exemplified for OLS and PLS regressions