Sample size effects in multivariate fitting of correlated data