Refine
Document Type
- Article (2)
Language
- English (2)
Has Fulltext
- yes (2)
Keywords
- Germany (1)
- cure survival models (1)
- fertility (1)
- parity progression (1)
- quantum (1)
- timing (1)
This article introduces double additive models to describe the effect of continuous covariates in cure survival models, thereby relaxing the traditional linearity assumption in the two regression parts. This class of models extends the classical event history models when an unknown proportion of the population under study will never experience the event of interest. They are used on data from the German Socio-Economic Panel (GSOEP) to examine how age at first birth relates to the timing and quantum of fertility for given education levels of the respondents. It is shown that the conditional probability of having further children decreases with the mother's age at first birth. While the effect of age at first birth in the third birth's probability model is fairly linear, this is not the case for the second child with an accelerating decline detected for women that had their first kid beyond age 30.
Fertility progression in Germany : An analysis using flexible nonparametric cure survival models
(2016)
Objective: This paper uses data from the German Socio-Economic Panel (GSOEP) to study the transition to second and third births. In particular, we seek to distinguish the factors that determine the timing of fertility from the factors that influence ultimate parity progression.
Methods: We employ cure survival models, a technique commonly used in epidemiological studies and in the statistical literature but only rarely applied to fertility research.
Results: We find that education has a different impact on the timing and the ultimate probability of having a second and a third birth. Furthermore, we show that the shape of the fertility schedule for the total population differs from that of ‘susceptible women’ (i.e., those who have a second or a third child).
Conclusions: Standard event history models conflate timing and quantum effects. Our approach overcomes this shortcoming. It estimates separate parameters for the hazard rate of having a next child for the ‘susceptible population’ and the ultimate probability of having another child for the entire population at risk.
Contribution: We go beyond standard cure survival models, also known as split population models, used in fertility research by specifying a flexible non-parametric model using Bayesian P-splines for the latent distribution (related to the timing of an extra birth) instead of a parametric model. Our approach is, so far, limited to time-constant covariates, but can be extended to include time-varying covariates as well.