USMDB Methods
Methods to construct the lifetable series for the US as a whole, the Census Regions and Divisions, the states and DC
The methods implemented to construct the lifetable series for the country as a whole, for the 4 Census Regions, 9 Census Divisions, all 50 states and the District of Columbia are strictly those of the Human Mortality Database and they are classic demographic methods.
The process for computing mortality rates and life tables can be described in five steps. All details are provided in the HMD Methods Protocol. Here is an overview of the process:
- Deaths. Death counts are collected at the finest level of detail available. Through our Data User Agreement with the NCHS, we have been able to obtain death records with the exact dates of birth and death. We can thus cross-tabulate deaths by year of birth and age at last birthday, i.e. by Lexis triangle. Note, however, that we have to aggregate death counts by single year of age or (at very high ages for instance) by five-year age group to avoid cell counts below 10 per our DUA with the NCHS. If raw data are aggregated, uniform methods are used to estimate death counts by completed age (i.e., age-last-birthday at time of death), calendar year of death, and calendar year of birth.
- Population size. Annual estimates of population size on January 1st are derived from the July 1st annual population estimates provided by the Census bureau or (when unavailable – e.g. for years before 1970 at the state level) by census data plus birth and death counts.
- Exposure-to-risk. Estimates of the population exposed to the risk of death during some age-time interval are based on annual (January 1st) population estimates, with a small correction that reflects the timing of deaths within the interval.
- Death rates. Death rates are always a ratio of the death count for a given age-time interval divided by an estimate of the exposure-to-risk in the same interval.
- Life tables. To build a life table, probabilities of death are computed from death rates. These probabilities are used to construct life tables, which include life expectancies and other useful indicators of mortality and longevity.
Methods to construct the lifetable series at the county level
Due to issues specific to small populations (large year-to-year fluctuations that obscure the true underlying mortality patterns) and to issues of confidentiality, the county-level lifetables are not constructed using classic demographic techniques, but a statistical model derived from Bayesian inference. Additional adjustments to the data are described in the USMDB Methods Protocol for the county series. The model is motivated and described in further detail in Dharamshi, Alexander, Winant and Barbieri, 2025. This is a modified version of the initial model described in Alexander, Zagheni and Barbieri, 2017. The initial model, which yielded estimates for both sexes combined, was modified to account for variations between the male and female mortality schedules within each county.
Note that the method only produces robust mortality estimates for populations of at least 10,000 people (of both sexes). Given that the median population size of the 3,144 US counties is 25,000, we had to aggregate the smallest counties into geographically adjacent groups within each state to reach the 10,000 population threshold where necessary. The lifetables were thus constructed for 2,330 single counties and 401 county aggregates.
The basic idea behind our chosen statistical approach is to fit observed deaths and population counts to a model of the underlying mortality rates. The model builds on characteristic age patterns in mortality curves, which are constructed using principal components from a set of reference mortality curves. These principal components create an underlying structure of the model in which many different kinds of shapes of mortality curves can be expressed. Withing a Bayesian hierarchical framework, information on mortality rates is then pooled across geographic space and smoothed over time. Geographic pooling allows mortality in smaller areas, where the underlying pattern are unclear, to be partially informed by mortality patterns in larger areas. A temporal smoothing component ensures that the parameters governing the shape of the mortality curve change gradually and in a relatively regular pattern over time. We add a constraint to the model so that county-level mortality rates, when aggregated to the state or national level, are consistent with the mortality rates observed at the aggregate level. The model is fitted in a Bayesian framework, with samples from the posterior distributions of the parameters obtained via a Markov Chain Monte Carlo (MCMC) algorithm.
Contact
If you have comments or questions about the database or if you encounter difficulties accessing the data, please reach out to us by emailing usmdb@mortality.org.