The impact of record-linkage bias in the Cox model
Version of Record online: 12 FEB 2010
© 2010 The Authors. Journal compilation © 2010 Blackwell Publishing Ltd
Journal of Evaluation in Clinical Practice
Volume 16, Issue 1, pages 92–96, February 2010
How to Cite
Baldi, I., Ponti, A., Zanetti, R., Ciccone, G., Merletti, F. and Gregori, D. (2010), The impact of record-linkage bias in the Cox model. Journal of Evaluation in Clinical Practice, 16: 92–96. doi: 10.1111/j.1365-2753.2009.01119.x
- Issue online: 12 FEB 2010
- Version of Record online: 12 FEB 2010
- Accepted for publication:22 August 2008
- breast cancer;
- Cox model;
- estimation bias;
- proportional hazard;
Rationale, aims and objectives Record linkage (RL) has become increasingly useful in health care administration, demographic studies, provision of health statistics and medical research. Linkage failure may occur when databases are affected by missing or inaccurate information. In particular, if the subsets of those who are not linked are not representative of the original population, the results obtained from linked data may be biased. This paper discusses the impact of incomplete RL on survival analysis.
Methods In our study we assess by simulations the potential impact of such bias, that we will refer to as RL, on the effect of the covariates in the Cox regression model. We also evaluate the RL bias introduced by an incomplete linkage procedure on the analysis of survival in a cohort of patients with breast cancer.
Results Our simulation study shows that the relative bias of the factors, which the linking probability depends on, reaches the threshold of 20%, and is never less than 5%. The bias observed in the simulation for a comparable scenario is consistent with the actual one estimated from the breast cancer records.
Conclusions Incomplete RL is rarely explicitly taken into account in the models for survival analysis. This study indicates that such a practice is potentially leading to inefficient and biased results, in particular in presence of medium or small sample sizes.