Regression models for censored serological data.

The impact was assessed of censored serological measurements on regression equations fitted to data from panels of sera tested by different laboratories, for the purpose of standardizing serosurvey results to common units. Several methods that adjust for censoring were compared, such as deletion, simple substitution, multiple imputation and censored regression. Simulations were generated from different scenarios for varying proportions of data censored. The scenarios were based on serological panel comparisons tested by different national laboratories and assays as part of the European Sero-Epidemiology Network 2 project. The results showed that the simple substitution and deletion methods worked reasonably well for low proportions of data censored (<20 %). However, in general, the censored regression method gave estimates closer to the truth than the other methods examined under different scenarios, such as types of equations used and violation of regression assumptions. Interval-censored regression produced the least biased estimates for assay data resulting from dilution series. Censored regression produced the least biased estimates in comparison with the other methods examined. Moreover, the results suggest using interval-censored regression methods for assay data resulting from dilution series.