Bias in correlations from selected samples of relatives: The effects of soft selection