This paper describes the steps involved in harmonizing a statistical method for the classification of illicit heroin using manufacturing impurity data. Sixteen linked samples were prepared from each of the five unrelated heroin seizures. Five links, totaling 80 samples were subjected to two sample weight approaches for liquid–liquid extraction. A wide range of intra-batch variations were observed in the dataset and this was probably due to the poor extraction efficiency and weight differences. Notwithstanding this, this paper aims to find the best clustering technique (pretreatment and clustering tool) that is able to minimize the intra-batch variation without undermining the inter-batch variation for sample classification. Three pretreatment methods were first evaluated by principle component analysis (PCA) and normalization coupled with standardization was found to be able to show promising results. Subsequently, hierarchical cluster analysis (HCA) was chosen to evaluate the pretreated data as this technique can clearly depict the relationships between samples on a dendogram. In the HCA, Ward-Squared Euclidean and Ward-Squared Pearson were found to be the successful linkage-distance combinations for clustering all the linked samples under five distinct groups.