Verbal autopsy coding: are multiple coders better than one?

Abstract

OBJECTIVE: To assess the impact on the reported cause-of-death patterns of a verbal autopsy coding strategy based on a review of every death by multiple coders versus a single coder. METHODS: Deaths in 45 villages (total population 180 162) in southern India were documented during 12 months in 2003-2004, and a standard verbal autopsy questionnaire was completed for each death. Two physician coders, each unaware of the other's decisions, assigned an underlying cause of death in accordance with the causes listed in the chapter headings of the International classification of diseases and related health problems, 10th revision (ICD-10). For the three chapter headings that applied to more than 100 of the deaths, agreement for subsets of causes of death within the chapter was also analysed. In the event of discrepancies, a third coder was used to finalize a cause of death. Cohen's kappa statistic (K) was used to measure levels of agreement between the two physician coders. FINDINGS: In total, 1354 deaths were documented, and a verbal autopsy was completed for 1329 (98%) of them. At the chapter heading level of the ICD-10, physician coders assigned the same cause to 1255 deaths (94%) (K = 0.93; 95% confidence interval: 0.92-0.94). The patterns of death derived from the causes assigned by each physician were all very similar to the patterns obtained through the consensus process, with the rank order of the 10 leading causes of death being the same for all three coding methods. CONCLUSION: Duplicate coding of verbal autopsy results has little advantage over a single-coder system for mortality surveillance or for identifying population patterns of death. Resources could be better diverted to other parts of the mortality surveillance process, such as validation.