Learning Individual Reproductive Behavior from Aggregate Fertility Rates via Neural Posterior Estimation
Daniel Ciganda, Ignacio Campón, Iñaki Permanyer, Jakob H Macke · 2025
Read Learning Individual Reproductive Behavior from Aggregate Fertility Rates via Neural Posterior Estimation by Daniel Ciganda, Ignacio Campón, Iñaki Permanyer, Jakob H Macke (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-22607.txt.
Introduction
Ciganda and co-authors tackle a gap in population research: birth rates by age are plentiful, but they hide the personal choices that drive them. The team couples a simulation model of individual reproductive life with a neural method that infers the model's settings from data alone. The model covers desired family size, timing, and contraceptive failure. Fitted only to aggregate fertility rates, it recovered the core behavioral parameters for cohorts in four countries.
More striking, the fitted model predicted facts it never saw, such as age at first sex and gaps between births, at the individual level. Because the method outputs complete synthetic life histories, it can feed forecasting models that usually demand rich survey data. For demographers, the appeal is behavioral insight drawn from the most widely available fertility records in the world.
Catalog record
| CREATOR | Daniel Ciganda, Ignacio Campón, Iñaki Permanyer, Jakob H Macke |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-22607
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning Individual Reproductive Behavior from Aggregate Fertility Rates via Neural Posterior Estimation Daniel Ciganda, Ignacio Campón, Iñaki Permanyer, Jakob H Macke Abstract While age-specific fertility rates (ASFRs) provide the most extensive record of reproductive change, their aggregate nature masks the underlying behavioral mechanisms that ultimately drive fertility trends. To recover these mechanisms, we develop a likelihood-free Bayesian framework that couples an individual-level model of the reproductive process with Sequential Neural Posterior Estimation (SNPE). This allows us to infer eight behavioral and biological parameters from just two aggregate series: ASFRs and the age-profile of planned versus unplanned births. Applied to U.S. National Survey of Family Growth cohorts and to Demographic and Health Survey cohorts from Colombia, the Dominican Republic, and Peru, the method reproduces observed fertility schedules and, critically, predicts out-of-sample micro-level distributions of age at first sex, inter-birth intervals, and family-size ideals, none of which inform the estimation step. Because the fitted model yields complete synthetic life histories, it enables behaviorally explicit population forecasts and supports the construction of demographic digital twins. Keywords— simulation-based inference, individual-level modelling, age-specific fertility rates, reproductive behavior, microsimulation 1 Introduction Globally declining fertility rates are driving a profound demographic transformation, presenting first-order societal challenges worldwide and underscoring the critical need for robust models to understand and anticipate future fertility trends. Age-specific fertility rates (ASFRs) —births per woman at each age— remain the workhorse input for these tasks because the underlying birth counts are collected continuously in almost every country’s vital-statistics system. ASFRs series therefore provide unparalleled temporal and geographic coverage. Yet ASFRs are only marginal aggregates. They reveal when births occur but not why . A long tradition of parametric models has been developed to describe ASFRs curves, smooth observed schedules, extrapolate fitted parameters, or convert projected total fertility rates into age patterns for cohort-component population forecasts ( Coale and Trussell 1974 ; Brass 1974 ; Schmertmann 2003 ; Chandola et al. 1999 ) . While existing approaches can excel at capturing aggregate ASFRs shapes, their parameters are typically macro-level abstractions. As such, they offer limited or elusive insight into the individual behavioural mechanisms driving fertility decisions ( Hoem and Hoem 1989 ) . Consequently, forecasts based on these models often cannot explicitly account for how evolving individual choices regarding the timing of parenthood, contraceptive practice, and birth-spacing and stopping decisions shape future fertility trends. This disconnect is striking in light of the explosion of micro - level fertility research based on rich individual data sources. Studies using birth histories, time-use diaries, and linked administrative records have uncovered nuanced relationships between education, labor market trajectories, union stability, and childbearing. Yet virtually none of that behavioural insight is incorporated into fertility projection models. The challenge of understanding contemporary fertility lies in bridging this micro-macro divide. This paper proposes a step toward this goal. We introduce a demographically interpretable individual-level stochastic model that captures key features of modern reproductive behavior, including stopping, spacing, and contraceptive failure. Estimating such a richly parameterised model from coarse data is a non-trivial inverse problem: many combinations of micro parameters can generate similar ASFRs curves. We show that adding the age profile of planned versus unplanned births to the set of ASFRs, provides enough information to identify the key parameters of this model. We fit the model to U.S. National Survey of Family Growth cohorts and to Demographic and Health Survey cohorts from Colombia, the Dominican Republic, and Peru, representing four settings that span markedly different fertility regimes. Using Sequential Neural Posterior Estimation (SNPE), a likelihood-free method that learns to approximate posterior parameter distributions from simulator outputs, we recover full posterior distributions of behaviourally meaningful parameters. Posterior predictive checks further confirm that the inferred parameters enable the simulator to accurately reproduce the observed aggregate fertility and unplanned fertility rates. Most significantly, we conduct out-of-sample validation, demonstrating that the parameters inferred solely from these aggregate population rates can successfully predict distributions of individual-level characteristics, such as the empirical distributions of age at first sexual intercourse, desired family sizes reported in surveys, and realized birth intervals. This capacity to predict micro-level outcomes not directly used in the estimation process provides strong evidence for the model’s structural validity and the inference procedure’s power. By establishing a demographically interpretable and statistically robust link between individual reproductive behaviors and population-level fertility outcomes, our framework offers a promising path towards more nuanced understanding and potentially more accurate and behaviorally grounded fertility forecasts. 1.1 Previous Work Efforts to model the reproductive process mechanistically have a long and distinguished tradition in demography. Pioneered by early work on fecundability ( Gini 1924 ) and significantly advanced by work on ’natural fertility’ ( Henry 1953 ) , this foundation led to a first wave of mathematical and early simulation models of the reproductive process ( Brass 1958 ; Singh 1963 ; Ridley and Sheps 1966 ; Potter 1972 ; Bongaarts 1977 , among others) . These models sought to capture the biological and behavioral drivers of family building with increasing realism. However, this vibrant tradition of individual-level modeling did not continue to evolve within demography at the same pace or in the same manner as in other scientific fields. While empirical fertility research using micro-data flourished ( Xie 2000 ) , the further development of these process-based simulators saw less sustained momentum. More critically, this stream of work remained largely disconnected from the concurrent revolution in statistical methodology for simulation-based inference. Techniques like Approximate Bayesian Computation (ABC) or other simulation-based inference (SBI) approaches that became influential in fields like ecology and genetics ( Beaumont 2010 ; Hartig et al. 2011 ) , were not systematically integrated in demography. This methodological lag created a persistent gap in formally linking mechanistic understanding to empirical observations in fertility research. Our own previous work showed that such a bridge is feasible in natural-fertility contexts —historical or religious communities without deliberate birth control— where a handful of biological parameters govern the reproductive process and can be inferred from ASFRs alone ( Ciganda and Todd 2024 ) . The present paper tackles the far more ambitious, and policy-relevant task, of modelling contemporary fertility, where stopping, spacing, and imperfect contraception introduce complex decision making and heterogeneity in reproductive dynamics. 1.2 Organization of the Article The remainder of this paper is structured as follows: Section 2 describes our individual-level computational model of the reproductive process. Section 3 details the data sources and the construction of ASFRs and Age-specific unplanned fertility rates (ASUFRs). Section 4 outlines the SNPE methodology. The Results section presents our findings on parameter identifiability via cross-validation, model fit through posterior predictive checks, and the out-of-sample validation against micro-level data. Finally, we discuss the implications of our findings, limitations, and avenues for future research in the Discussion. 2 Model The transition from a natural to a modern fertility regime in human populations is marked by the emergence of stopping and spacing behaviors (deciding whether and when to have children), representing a shift from a reproductive process largely determined by physiological factors to one shaped by individual preferences. As a result, fertility patterns become more diverse, with family formation increasingly tailored to individual or couple-level choices rather than biological constraints alone. Building on these insights, our model represents the