Joshua R. Goldstein’s research while affiliated with University of California, Berkeley and other places

Behind the steady march of progress toward longer life expectancy in many low‐mortality countries, there have been setbacks even before the Covid‐19 pandemic. In this paper, we use an exploratory approach to describe the temporal structure, age patterns, and geographic aspects of life expectancy reversals. We find that drops in life expectancy are often followed by larger than average improvements, which tells us that most reversals are transitory with little long‐term influence. The age structure of mortality decline when life expectancy falls is tilted toward older ages, a pattern that is quite different from the general pattern of mortality improvement. Geographic analysis shows that mortality reversals are correlated across neighboring countries like Italy and France, or Canada and the United States. These findings are consistent with contagious disease and weather being important causes of life expectancy reversals. We conclude with a discussion of implications for formal modeling and forecasting of mortality to accommodate these patterns that violate some standard assumptions.

When will the human population peak? In this article, we build on classical results by Ansley Coale, who showed that when fertility declines steadily, births reach their maximum before fertility reaches replacement level, and the decline in total population size does not occur until several decades after fertility has reached that level. We extend Coale's results by modeling longevity increases, net immigration, and a slowdown in fertility decline that resembles current projections. With these extensions, our models predict a typical lag between replacement-level fertility and population decline of about 35 to 40 years, consistent with projections by the United Nations and about 15 years longer than the lag predicted by Coale. Our analysis helps reveal underlying factors in the timing of peak population.

In the United States, much has been learned about the determinants of longevity from survey data and aggregated tabulations. However, the lack of large-scale, individual-level administrative mortality records has proven to be a barrier to further progress. We introduce the CenSoc datasets, which link the complete-count 1940 U.S. Census to Social Security mortality records. These datasets—CenSoc-DMF (N = 4.7 million) and CenSoc-Numident (N = 7.0 million)—primarily cover deaths among individuals aged 65 and older. The size and richness of CenSoc allows investigators to make new discoveries into geographic, racial, and class-based disparities in old-age mortality in the United States. This article gives an overview of the technical steps taken to construct these datasets, validates them using external aggregate mortality data, and discusses best practices for working with these datasets. The CenSoc datasets are publicly available, enabling new avenues of research into the determinants of mortality disparities in the United States.

New advances in data linkage provide mortality researchers with access to administrative datasets with millions of mortality records and rich demographic covariates. Although these new datasets allow for high-resolution mortality research, administrative mortality records often have technical limitations, such as limited mortality coverage windows and incomplete observation of survivors. We describe a method for fitting truncated distributions that can be used for estimating mortality differentials in administrative data. We apply this method to the CenSoc datasets, which link the United States 1940 Census records to Social Security administrative mortality records. Our approach may be useful in other contexts where administrative data on deaths are available. As a companion to the paper, we release the R package gompertztrunc, which implements the methods introduced in this paper.

Can the names parents gave their children give us insights into how parents in historical times planned their families? In this study, we explore whether the names given to the firstborn child can be used as indicators of family-size preferences and, if so, what this reveals about the emergence of intentional family planning over the course of the demographic transition. We analyze historical populations from 1850 to 1940 in the United States, where early fertility control and large sample sizes allow separate analyses of the White and Black populations. We also analyze Norway from 1800 to 1910, where there was a much later fertility transition. A split-sample method allows automated scoring of each name in terms of predicted family size. We find a strong relationship between naming and family size in the U.S. White population as early as 1850, for the Black population beginning in 1940, and for the Norwegian population in 1910. These results provide new evidence of the emergence of "conscious calculation" during the fertility transition. Our methods may also be applicable to modern high-fertility populations in the midst of fertility decline.

New advances in data linkage provide mortality researchers with access to administrative datasets with millions of mortality records and rich demographic covariates. Although these new datasets allow for high-resolution mortality research, administrative mortality records often have technical limitations, such as limited mortality coverage windows and incomplete observation of survivors. We describe a method for fitting truncated distributions that can be used for estimating mortality differentials in administrative data. We apply this method to the CenSoc datasets, which link U.S. 1940 Census records to Social Security administrative mortality records. Our approach may be useful in other contexts where administrative data on deaths are available. As a companion to the paper, we release the R package gompertztrunc, which implements the methods introduced in this paper.

Background: While much progress has been made in understanding the demographic determinants of mortality in the United States using individual survey data and aggregate tabulations, the lack of population-level register data is a barrier to further advances in mortality research. With the release of Social Security application (SS-5), claim, and death records, the National Archives and Records Administration (NARA) has created a new administrative data resource for researchers studying mortality. We introduce the Berkeley Unified Numident Mortality Database (BUNMD), a cleaned and harmonized version of these records. This publicly available dataset provides researchers access to over 49 million individual-level mortality records with demographic covariates and fine geographic detail, allowing for high-resolution mortality research. Objective: The purpose of this paper is to describe the BUNMD, discuss statistical methods for estimating mortality differentials based on this deaths-only dataset, and provide case studies illustrating the high-resolution mortality research possible with the BUNMD. Methods: We provide detailed information on our procedure for constructing the BUNMD dataset from the most informative parts of the publicly available Social Security Numident application, claim, and death records. Contribution: The BUNMD is now publicly available, and we anticipate these data will facilitate new avenues of research into the determinants of mortality disparities in the United States.

While much progress has been made in understanding the demographic determinants of mortality in the United States using individual survey data and aggregate tabulations, the lack of population-level register data is a barrier to further advances in mortality research. With the release of Social Security application (SS-5), claim, and death records, the National Archives and Records Administration (NARA) has created a new administrative data resource for researchers studying mortality. We introduce the Berkeley Unified Numident Mortality Database (BUNMD), a cleaned and harmonized version of these records. This publicly available dataset provides researchers access to over 49 million individual-level mortality records with demographic covariates and fine geographic detail, allowing for high-resolution mortality research.

BACKGROUND The criteria used to allocate scarce COVID-19 vaccines are hotly contested. While some are pushing just to get vaccines into arms as quickly as possible, others advocate prioritization in terms of risk. OBJECTIVE Our aim is to use demographic models to show the enormous potential of vaccine risk-prioritization in saving lives. METHODS We develop a simple mathematical model that accounts for the age distribution of the population and of COVID-19 mortality. This model considers only the direct live-savings for those who receive the vaccine, and does not account for possible indirect effects of vaccination. We apply this model to the United States, Japan, and Bangladesh. RESULTS In the United States, we find age-prioritization would reduce deaths during a vaccine campaign by about 93 percent relative to no vaccine and 85 percent relative to age-neutral vaccine distribution. In countries with younger age structures, such as Bangladesh, the benefits of age-prioritization are even greater. CONTRIBUTION For policy makers, our findings give additional support to risk-prioritized allocation of COVID-19 vaccines. For demographers, our results show how the age-structures of the population and of disease mortality combine into an expression of risk concentration that shows the benefits of prioritized allocation. This measure can also be used to study the effects of prioritizing other dimensions of risk such as underlying health conditions.