Agnieszka Werpachowska - Applicability of Neural Networks in Microsimulations: A Case Study on Fertility Analysis

  • Presenting author: Agnieszka Werpachowska (Averisera Ltd)

  • Authors: A. Werpachowska

  • Session: B03B - Population [3] - Tuesday 14:00-15:30 - Marietta-Blau Hall

  • Slides: PDF

Microsimulations offer valuable insights into the dynamics of populations, drawing on well-understood relationships between variables representing key population characteristics and individual attributes. However, sometimes these relationships can be very complicated or contain subtle interactions and dependencies that may not be easily captured by traditional analytical approaches or domain-specific knowledge but remain crucial in modeling intricate dynamics within microsimulation models. A potential approach to handle such cases is using neural networks, which have emerged as powerful tools for modeling complex trends and patterns in various domains. The flexibility and capacity of neural networks to learn hidden patterns from high-dimensional data offer microsimulation researchers additional insight into the complex and noisy information. The generalizability and transferability of trained models enable them to utilize learned representations to make predictions and analyze scenarios or interventions based on modified new inputs. This work explores the potential of neural networks as a tool to enhance microsimulations by analyzing the factors influencing fertility rates in the UK. These factors include population age structure, economic indicators (such as unemployment rate or gender pay gap), female education attainment, emigration, and family-friendly policies. We incorporate them as additional features to a recurrent neural network model, RNN [1], aiming to uncover their influence on the modelled trend. Contrary to popular belief, neural networks are not black-box models. Various techniques enable us to assess the importance and contributions of different features. Visualization of neuron activations shows to which features the network responds, revealing which factors influence the fertility rate. Examining the values of weights assigned to neural connections reveals the strength of this influence. Gradient-based feature attribution methods work by computing the gradients of the output with respect to the input features, reflecting the strength and direction of their impact. A standard sensitivity analysis can also be performed by perturbing the input features and observing the resulting changes in the model’s outcome. We apply these methods to understand the fertility trends and forecast different scenarios and interventions. Additionally, we propose a simple method of estimating the prediction interval for the presented results.

[1] A. Werpachowska, “Forecasting the impact of state pension reforms in post-Brexit England and Wales using microsimulation and deep learning”, Proceedings of PenCon 2018 Pensions Conference, 2018