The progress in microarray technology is evident and huge amounts of gene expression data are currently being produced. A complicating matter is that there are various sources of uncertainty in microarray experiments, as well as in the analysis of expression data. This problem has generated an increased interest in the validation of methods for analysis of expression data. Clustering algorithms have been found particularly useful for the study of coexpressed genes, and this paper therefore concerns the robustness of partitional clustering algorithms. These algorithms use a predefined number of clusters and assign each gene to exactly one cluster. The effect of repeated clustering using identical algorithm parameters and input data is investigated for the self-organizing map (SOM) and the $k$-means algorithm. The susceptibility to measurement noise is also studied. A reproducibility measure is proposed and used to assess the results from the performed clustering experiments. Well-known publicly available datasets are used. Results show that clusterings are not necessarily reproducible even when identical algorithm parameters are used, and that the problems are aggravated when measurement noise is introduced.