Dissertations, Theses, and Capstone Projects

Date of Degree

6-2026

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy

Program

Economics

Advisor

Lilia Maliar

Committee Members

Frank Heiland

George Vachadze

Subject Categories

Economics | Labor Economics | Other Economics

Keywords

Aging; Retirement; Social Security; Machine Learning; Dynamic Discrete Choice; Pensions and Job Mobility

Abstract

Population aging has made the study of older workers’ economic behavior increasingly important for understanding labor markets, retirement systems, and public policy. In many advanced economies, rising old-age dependency ratios have increased the need to understand labor-supply, retirement, and savings decisions later in life (OECD, 2025). In the United States, these issues are especially important because Social Security faces long-run financing pressure, with the Old-Age and Survivors Insurance Trust Fund projected to be depleted around 2033 (U.S. Social Security Administration, Office of the Chief Actuary, 2025). Longer life expectancy increases the expected length of retirement and raises the fiscal importance of claiming decisions and labor-force participation at older ages. At the same time, the ability of older workers to remain employed, move to better jobs, or adjust the timing of retirement depends on the institutional structure of labor markets, pensions, health insurance, and public retirement programs. Understanding these decisions is therefore important not only for individual welfare, but also for the design of public policy in aging economies.

This dissertation studies late-career economic decisions, including retirement, Social Security claiming, job mobility, and consumption-saving choices, using three complementary approaches. The first chapter asks whether modern machine-learning methods can improve predictions of retirement and Social Security claiming. The second chapter develops a neural-network method for solving life-cycle models with discrete and continuous choices. The third chapter estimates a structural dynamic model of job switching and retirement to quantify pension-induced mobility frictions among older workers. In summary, the chapters show how data-driven prediction, computational methods, and structural estimation can be used jointly to understand the economic behavior of older households.

A central theme of the dissertation is that retirement is not a single isolated decision. It is closely connected to employment histories, job mobility, pension incentives, health insurance, and the structure of life-cycle choice. Chapter 1 begins from the prediction problem. Using rich panel data from the Health and Retirement Study, it shows that job-history and health-insurance variables contain substantial predictive information for retirement and Social Security claiming. This finding motivates a deeper economic question: why do employment histories matter so much for late-career decisions? Chapter 3 addresses one possible mechanism by estimating how pension coverage changes the cost of switching jobs. If pension plans, especially defined-benefit plans, make mobility costly, then past job mobility and current pension coverage can shape both retirement timing and the value of continued work.

Chapter 2 provides the computational bridge between the predictive and structural parts of the dissertation. Many retirement models require workers to choose among discrete alternatives, such as work or retirement, while also choosing continuous variables, such as consumption and savings. These models often generate kinks, thresholds, and discontinuities in policy functions. Standard numerical methods can become difficult to apply in richer environments with high-dimensional states, multiple discrete choices, aggregate uncertainty, or equilibrium feedback. Chapter 2 proposes a discontinuity-aware neural-network architecture, JumpBlock, as a step toward scalable solution methods for such models. Although Chapter 3 is estimated using a dynamic discrete-choice framework without continuous savings, the method developed in Chapter 2 points toward a natural extension: a richer model in which older workers jointly choose job mobility, retirement, consumption, and savings.

The first chapter, joint with Lilia Maliar, studies retirement and Social Security claiming as prediction problems. Accurate forecasts of claiming behavior are important for evaluating the fiscal outlook of Social Security and for assessing the distributional consequences of policy reform. The chapter compares gradient boosted trees with the benchmark Modeling Income in the Near Term (MINT) model used by the Social Security Administration. The machine-learning model improves predictive accuracy for both retirement and Social Security claiming. In a two-step prediction exercise, the benchmark model overpredicts the beneficiary rate by more than three percentage points, while the gradient-boosted-tree model has an error below one percentage point. A conservative calculation implies that the difference corresponds to approximately $35.16 billion per year in aggregate benefits. The chapter also uses Shapley values to interpret the fitted model and shows that the contributions of important variables are often nonlinear. Beyond improving prediction, the results point to job history, health insurance, and household characteristics as important predictors of late-career decisions.

The second chapter develops a computational method for discrete-continuous dynamic choice models. It adapts the DeepVPD approach to a retirement-consumption model and introduces JumpBlock, a neural-network layer designed to approximate discontinuous or threshold-like policy behavior. The forward pass of the layer retains a discontinuous Heaviside-gated structure, while the backward pass uses a smooth surrogate gradient. This construction allows the network to represent sharp policy changes while still providing informative gradients during training. The method is evaluated in the retirement-consumption model of Iskhakov et al. (2017), where analytic or highly accurate benchmark solutions are available. Across deterministic and stochastic taste-shock specifications, JumpBlock approximates the benchmark policy functions closely and generally improves on a parameter-matched standard multilayer perceptron, especially near discontinuities. The chapter therefore contributes to the growing literature on neural-network solution methods for dynamic economic models and highlights the importance of architectures that reflect the economic structure of the problem.

The third chapter studies why late-career job-to-job mobility is low. Job changes can allow older workers to move into positions with better wages, greater flexibility, lower stress, or less physically demanding tasks. Yet mobility rates among older workers are much lower than among younger workers. The chapter focuses on pension coverage as a key friction. Defined-benefit pensions are often tenure-linked and backloaded, so leaving an employer can imply a loss of future pension accrual. Defined-contribution plans are more portable in principle, but may still be associated with vesting rules, compensation premia, search frictions, uncertainty, or inertia. To quantify these mechanisms, the chapter estimates a dynamic discrete-choice model in which workers choose whether to stay with the current employer, switch employers, or retire. The model allows switching costs to differ by pension type and includes unobserved heterogeneity using the framework of Arcidiacono and Miller (2011). The estimates show that pension coverage substantially raises switching costs, with the largest frictions associated with defined-benefit plans and smaller but still meaningful frictions associated with defined-contribution plans. Counterfactual simulations indicate that reducing pension-related switching costs increases job-to-job mobility and modestly delays retirement.

The three chapters contribute to different literatures, but they are unified by a common question: how should economists model late-career decisions when the relevant environment is rich, nonlinear, and dynamic? Chapter 1 shows that flexible prediction methods can improve prediction accuracy and reveal which variables contain information about retirement and claiming decisions. Chapter 2 develops a computational tool for solving dynamic models in which policy functions may be non-smooth because of discrete choices. Chapter 3 uses structural estimation to interpret one important mechanism suggested by the predictive evidence: pension-induced mobility frictions. In this sense, the dissertation moves from prediction, to computation, to structural interpretation.

The dissertation also illustrates the complementarity between machine learning and structural economics. Machine learning is useful for prediction and variable discovery, especially when the data contain many potentially relevant covariates and nonlinear interactions. Structural models are useful for interpreting behavior, estimating economic frictions, and performing counterfactual policy analysis. Computational methods connect these approaches by making it possible to solve richer models that more closely resemble the decisions households actually face. Rather than treating these approaches as substitutes, this dissertation uses them as complementary tools for studying aging, retirement, and late-career labor markets.

The remainder of the dissertation is organized as follows. Chapter 1 studies retirement and Social Security claiming using gradient boosted trees and Shapley values. Chapter 2 develops JumpBlock and applies it to a discrete-continuous retirement-consumption model. Chapter 3 estimates pension-induced switching costs in a dynamic discrete-choice model of older workers and evaluates counterfactual pension reforms.

Download

Share

COinS