Correlation estimation with bivariate censored data via alternating regressions
James, I. (2010) Correlation estimation with bivariate censored data via alternating regressions. In: ASC 2010 Australian Statistical Conference Statistics in the West: Understanding our World., 6 - 10 December, Fremantle, Western Australia.
Estimation of correlation when both variables are potentially right censored is confounded by identifiability issues and typically requires assumptions about the nature of the association. Semiparametric approaches have included the use of copulas, frailties or mixed models. Given that correlation is most meaningful within the context of linear association we consider iterative processes via alternating or 'criss-cross' censored linear regressions to estimate the bivariate correlation, under the assumption that each variable regresses linearly on the other. Two approaches are considered. In the first, value fragments corresponding to re-distributed censored responses from one regression are imputed as weighted explanatory variables in the alternative and the process iterated. Correlation is estimated via the two sets of slope parameters. The second replaces the weighted values by data augmentation. The efficacies of the approaches are compared and illustrated via simulations. Our initial results suggest that both methods compensate well for the censoring even with a significant proportion of cases having both variables censored, with the data augmentation approach slightly less biased. Additional (uncensored) covariates can be readily incorporated. We demonstrate the method via analysis of correlated immunological and virological measures on HIV-1 positive patients from the WAH IV Cohort, which may be incomplete for a variety of reasons.
|Publication Type:||Conference Item|
|Murdoch Affiliation:||Centre for Clinical Immunology and Biomedical Statistics|
|Item Control Page|