Forecasting Cohort Childlessness: Bayesian Modeling Based on Historical Patterns in the Human Fertility Database

Carl P. Schmertmann, Florida State University

We propose a new Bayesian approach to forecasting cohort childlessness. Combining historical and contemporary data from the Human Fertility Database (HFD), we estimate a posterior distribution for the Lexis surface of age-specific first-birth rates for US cohorts born 1950-1992. Past rates on this surface are known with high precision from the HFD, while future rates must be forecast. Our approach combines estimation of past and future rates in a single model, using historical HFD data to build priors and thus identify likely (and unlikely) age and time patterns across Lexis surfaces. The resulting forecast of first-birth rates and cohort childlessness automatically includes uncertainty estimates. Among many other results, our forecast indicates that US childlessness, which is currently falling slightly, will reach a minimum for women born in the 1970s, and will almost certainly be higher for those born in the 1980s.