Clustering with missing and left-censored data: A simulation study comparing multiple-imputation-based procedures

Cluster analysis, commonly used to explore large biomedical datasets, can be challenging, notably due to missing data or left-censored data induced by the sensitivity limits of the biochemical measurement method. Usually, complete-case analysis, simple imputation, or stochastic simple imputation are applied before clustering. More recently, consensus methods following multiple imputation have been proposed. However, they ignore left-censoring and do not allow the number of clusters to vary across the partitions of each imputed dataset. Here, we developed a consensus-based clustering algorithm in which left-censored data are taken into account using a modified multiple imputation method and the number of clusters is estimated for each imputed dataset. A simulation study was conducted to assess the performance in terms of the number of clusters, the percentage of unclassified observations, and the adjusted Rand index. The simulation results showed that the investigated method works well compared to several alternative approaches. A real-world application in breast cancer patients showed that the proposed method may reveal novel clusters of patients.