Hidden Group Time Profiles: Heterogeneous Drawdown Behaviours in Retirement

Igor Balnozan, Denzil G. Fiebig, Anthony Asher, Robert Kohn, Scott A. SissonarXiv 2020 · DOI · arXiv

Abstract

This article investigates retirement decumulation behaviours using the Grouped Fixed-Effects (GFE) estimator applied to Australian panel data on drawdowns from phased withdrawal retirement income products. Behaviours exhibited by the distinct latent groups identified suggest that retirees may adopt simple heuristics determining how they draw down their accumulated wealth. Two extensions to the original GFE methodology are proposed: a latent group label-matching procedure which broadens bootstrap inference to include the time profile estimates, and a modified estimation procedure for models with time-invariant additive fixed effects estimated using unbalanced data.

Introduction

Decumulation—the process of spending down assets during retirement—has been described as the "hardest problem in finance." For economists and policy-makers, modeling this phase is exceptionally difficult because retirees do not behave as a monolithic group. Traditional econometric models, such as the standard Two-Way Fixed-Effects (2WFE) approach, typically assume that while individuals may have different baseline spending levels, they respond to external shocks—like the Global Financial Crisis (GFC) or policy shifts—in a uniform way.

This paper argues that such "average" models are fundamentally misleading. If one group of retirees slashes spending during a crisis while another remains steady, the average result describes neither group accurately. By applying an enhanced Grouped Fixed-Effects (GFE) estimator to longitudinal data, this research identifies seven distinct latent groups of Australian retirees, revealing that most rely on simple behavioral heuristics rather than complex financial optimization.

Methodological Innovations

The researchers utilize the GFE estimator, a technique that functions like $k$-means clustering within a linear regression framework. It simultaneously assigns individuals to latent groups and estimates group-specific time profiles. To make this tool more robust for real-world applications, the paper introduces two significant methodological extensions:

  1. Label-Matching Procedure: In bootstrap simulations used to calculate statistical significance, group labels often "switch" between runs (e.g., Group A in one run might be identified as Group B in the next). The authors propose a permutation-based matching algorithm to fix these labels, allowing for valid statistical inference on the group time profiles themselves.
  2. Unbalanced Data Modification: Real-world panel data is rarely "balanced" (where every subject is observed at every time point). The authors modified the GFE estimator to ensure it remains consistent with standard econometric models even when data is missing for certain periods, a common occurrence in administrative datasets.

Data and Context

The study analyzes a high-quality Australian panel dataset of 9,516 individuals over a 12-year window (2004–2015). The data tracks drawdowns from Account-Based Pensions (ABPs), which are popular phased-withdrawal products in Australia. This period is particularly significant as it covers the market volatility of the GFC and several major changes to legislated minimum drawdown rates, providing a natural experiment to observe how different groups respond to external pressure.

Key Insights: The Seven Faces of Retirement

The analysis identifies $G=7$ as the optimal number of behavioral groups. The findings suggest that the "average" retiree is a myth; instead, retirees tend to follow specific "rules of thumb":

  • The Constant Dollar Crowd: Approximately two-thirds of the sample belong to groups that attempt to keep their nominal dollar withdrawals constant, regardless of how their account balances fluctuate.
  • The Policy Followers: Roughly 14% of the sample (Group 5) follow the government’s legislated minimum drawdown rates almost perfectly. This group showed a "step-like" pattern in their spending, closely mirroring regulatory changes.
  • The Engaged Adjusters: Group 4 appeared to be highly "engaged," constantly tuning their drawdown rates in response to market conditions until 2011, after which they significantly reduced their spending to preserve capital.

Results and Policy Implications

Moving from a single-group model to a multi-group model significantly changes the understanding of retirement spending. For instance, the multi-group model revealed that a 10% increase in account balance was associated with a roughly 1.46% decrease in the rate of drawdown—a nuance missed by simpler models that suggest a more direct, positive correlation between wealth and spending.

Furthermore, responsiveness to policy is highly concentrated. While Group 5 slashed drawdowns by 50% when the government introduced concessional minimum rates during the GFC, other groups barely reacted, continuing their "constant dollar" heuristic. This suggests that policy levers intended to influence retiree spending may only affect a small subset of the population.

Conclusion

This research demonstrates that retirement decumulation is characterized by deep heterogeneity. By providing new tools to identify and validate latent group behaviors, the authors offer a more accurate framework for analyzing how retirees actually manage their wealth. Future work may apply this label-matching GFE framework to other areas of behavioral microeconomics, such as labor supply or household consumption, where unobserved group-level responses to shocks are likely to exist.

Summary updated

Abstract

This article investigates retirement decumulation behaviours using the Grouped Fixed-Effects (GFE) estimator applied to Australian panel data on drawdowns from phased withdrawal retirement income products. Behaviours exhibited by the distinct latent groups identified suggest that retirees may adopt simple heuristics determining how they draw down their accumulated wealth. Two extensions to the original GFE methodology are proposed: a latent group label-matching procedure which broadens bootstrap inference to include the time profile estimates, and a modified estimation procedure for models with time-invariant additive fixed effects estimated using unbalanced data.
BibTeX
@misc{balnozan2020hidden,
  author = {Igor Balnozan and Denzil G. Fiebig and Anthony Asher and Robert Kohn and Scott A. Sisson},
  title = {{Hidden Group Time Profiles: Heterogeneous Drawdown Behaviours in Retirement}},
  year = {2020},
  eprint = {2009.01505},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2009.01505},
}
Figure