Identifying Statistical Bias in Dataset Replication