Improving the usefulness of US mortality data: new methods for reclassification of underlying cause of death

Mortality data are affected by miscertification of the medical cause of death deaths and changes to cause of death classification systems. We present both mappings of ICD9 and ICD10 to a unified list of causes, and a new statistical model for reducing the impact of misclassification of cause of death.

We propose a Bayesian mixed-effects multinomial logistic model that can be run on individual record level death certificates to reclassify "garbage-coded" deaths onto causes that are more meaningful for public health purposes. The model uses information on the contributing causes of death and demographic characteristics of each decedent to make informed predictions of the underlying cause of death. We apply our method to death certificate data in the US from 1979 to 2011, creating more directly comparable series of cause-specific mortality for 25 major causes of death.

We find that many death certificates coded to garbage codes contain other information that provides strong clues about the valid underlying cause of death. In particular, a plausible underlying cause often appears in the contributing causes of death, implying that it may be incorrect ordering of the causal chain and not missed cause assignment that leads to many garbage-coded deaths. We present an example that redistributes 48 % of heart failure deaths to other cardiovascular diseases, 25 % to ischemic heart disease, and 15 % to chronic respiratory diseases.

Our methods take advantage of more detailed micro-level data than is typically considered in garbage code redistribution algorithms, making it a useful tool in circumstances in which detailed death certificate data needs to be aggregated for public health purposes. We find that this method gives different redistribution results than commonly used methods that only consider population-level proportions.