Identification of influential observations based on binary particle swarm optimization in the cox PH model. Issue 3 (3rd March 2020)

• Record Type: Journal Article
• Authors: Sancar, Nuriye; Inan, Deniz
• Is Part Of: Communications in statistics. Volume 49:Issue 3(2020)
• Page Start: 567
• Page End: 590
• Publication Date: 2020-03-03
• Subjects: Cox PH Model -- Influential Observations -- Binary Particle Swarm Optimization -- Likelihood Displacement -- Masking Effect -- Swamping Effect; Mathematical statistics -- Periodicals; Mathematical statistics -- Data processing -- Periodicals; Digital computer simulation -- Periodicals; 519.5
• DOI: 10.1080/03610918.2019.1682156
• Languages: English
• ISSNs: 0361-0918

Abstract

Abstract: Proper identification of influential observations should be an integral and significant part of the Cox modeling process. This is because the failure to identify influential observations may have negative effect on the estimates acquired from the process. Furthermore, in survival analysis, influential observations frequently propose that the model has imperfections, including poorly specified factor, non-proportional hazards, loss of important information and/or a variable that has been omitted. Nevertheless, many procedures have been developed for the identification of a single influential observation based on the leave-one-out method. However, the results from leave-one-out diagnostic techniques are often misleading as a result of swamping and masking problems in the presence of multiple influential observations in the dataset. In this paper, identification of the optimal set of influential observations problem has been considered as the combinatorial optimization problem and a new simultaneous approach for identification of the optimal set of influential observations is proposed based on Binary Particle Swarm Optimization (BPSO) approach in the Cox PH model. The performance of the proposed BPSO-based approach and conventional diagnostic techniques have been compared according to various evaluation criteria by simulation studies. The performance of the BPSO-based approach has been also demonstrated by the clinical real dataset.