Abstract

Accurate forecasts of retirement and Social Security (SS) claiming behavior are important for projecting the long-run costs of public pension and social-insurance systems. We study whether machine learning methods, applied to a rich set of household-level variables, can improve predictions of retirement and SS claiming decisions. Our model predicts the share of SS beneficiaries with an error of less than 1 percentage point, while the benchmark model used by the Social Security Administration (SSA) has an error of more than 3 percentage points. A conservative calculation implies that this difference corresponds to about 35.16 billion dollars per year in aggregate benefits. In a 2020 forecasting exercise using models trained on pre-2018 data, our model retains strong predictive performance during COVID-19 and outperforms the SSA benchmark. We also find that our model performs better across groups defined by expected SS benefits, with especially large gains among individuals with high expected benefits. In addition, the variables selected by our model differ from those emphasized in the SSA benchmark. Using Shapley values, we show that variable contributions are often nonlinear.

Keyword

Abstract

We study external validity within the context of instrumental variable estimation. The key assumption we impose for external validity is conditional external unconfoundedness among compliers, which means that the treatment effect and target selection are independent among compliers conditional on covariates. We study this assumption by using a case study about the impact of solid fuel usage on women’s average cooking time. Among the six countries examined, we find no statistical evidence that the assumptions for external validity are violated for four countries (Ethiopia, Honduras, Kenya, and Zambia). Conversely, in Cambodia and Nepal, we find low external validity. These results provide suggestive evidence that the assumptions for external validity are violated for these two countries.

Keyword

Abstract

Using data from the Multi-Tier Framework Survey (MTF) conducted in Nepal, we explore how geographical factors, specifically land slope and elevation, impact the adoption of LPG stoves. We employ a logit model to analyze the factors influencing households’ choices regarding LPG stove adoption and generate slope and elevation data. Overall, we find that the estimates of the average slope are robust and statistically significant, but those for the average elevation are sensitive to model specifications and smaller compared to the slope. We find that the Kathmandu region is important in the analysis, and slope and elevation have nonlinear effects. Additionally, we show that geographical factors are similarly important across different household expenditure quintile groups, except for the lowest.

Keyword

Abstract

I study how pension coverage affects job mobility among older U.S. workers. Using panel data from the Health and Retirement Study and a dynamic discrete-choice model with unobserved heterogeneity, I estimate wage-equivalent switching costs for workers ages 50–75. In the model, I find that pension coverage increases switching frictions. In the main specification, defined-contribution coverage raises annualized switching costs by about $6,400 per worker-year (10 percent of the average wage) and defined-benefit coverage by about $8,600 (14 percent), in real 2024 dollars. Counterfactual simulations indicate that halving pension-related switching costs raises job-to-job mobility among DB holders by roughly 34 to 41 percent, with implied welfare gains of roughly $44 to $462 per worker-year across the proportional reforms and specifications, concentrated among pension-covered workers. The robust finding is that pension coverage of either type carries large switching costs: both the DB and the DC coefficients are statistically significant in every specification, so pension coverage is an important source of late-career labor-market frictions.

Keyword

Abstract

Dynamic models with discrete choice are widely used in economics, but empirical applications remain challenging because discrete actions generate non-convexities and non-smooth policy rules, and rich state vectors make grid-based dynamic programming computationally costly. This survey reviews recent computational advances for solving and estimating dynamic discrete choice and dynamic discrete-continuous choice models. We organize the survey around three common computational bottlenecks: high-dimensional state spaces, non-smooth decision boundaries, and fixed-point problems in Bellman equations and equilibrium conditions. On the solution side, we discuss scalable benchmarks and neural-network-based methods that approximate policy functions and choice probabilities. On the estimation side, we summarize full-solution and two-step conditional-choice-probability approaches, and we emphasize developments that mitigate the nested fixed-point problem, including mathematical programming with equilibrium constraints, Euler-type formulations, continuous-time reformulations, sieve approximation, and machine-learning procedures.

Keyword

Abstract

We study the impact of policy loan programs on the productivity of small and medium enterprises (SME) in South Korea from 2013 to 2022. Using a robust Difference-in-Differences estimator for average treatment effects (ATE) and a Generalized Random Forest for conditional ATE, we focus on two productivity measures: total factor productivity (TFP) and its persistent component. We find that subsidies have a positive and increasing, but delayed, effect on TFP. Meanwhile, we find a positive and increasing, but not significant, effect on the persistent component of productivity, which warrants a cautious interpretation of the policy's ability to encourage sustainable long-run growth. We also find significant heterogeneity: TFP gains are greatest for financially constrained firms, which are usually younger and smaller firms. The heterogeneity in persistent productivity is even more pronounced, suggesting that a "one-size-fits-all" approach is insufficient for long-term impact. Instead, targeted support is necessary to promote sustainable growth among SMEs.We evaluate policy loans to Korean SMEs using confidential administrative data for 2013-2022. Combining propensity-score matching with staggered difference-in-differences estimation, formal sensitivity analysis for parallel-trends violations, and generalized random forests for treatment-effect heterogeneity, we find that revenue-based total factor productivity rises with a delay, by 5.5% two years and 6.4% three years after loan receipt. Gains are largest among young, small, less leveraged, and low-productivity firms. Effects on the persistent component of productivity are imprecisely estimated and sensitive to sample composition, so the evidence neither confirms nor rules out lasting productivity improvements.

Keyword

Abstract

In this paper, I approximate the solution of a discrete-continuous dynamic choice model using deep neural networks. The model of interest is an overlapping generation (OLG) model in which the individual needs to make two decisions: how much to consume, and when to retire. Overall, the deep neural networks designed in this paper approximate the analytic solution of consumption and the retirement probability relatively well. However, some points are relatively far from the analytic solution at the kink points and the discontinuity points, especially for the individual of age 1.

Keyword

Abstract

Empirical evidence from instrumental variable (IV) studies often guides policy decisions beyond the original study setting. Yet IV estimates identify local average treatment effects (LATEs) that may not generalize when the composition of compliers differs across populations. This paper examines how machine learning methods can improve the external validity of IV estimates. Using an empirical application on the effect of solid fuel use on cooking time across six developing countries and a series of simulation experiments, we compare the benchmark interacted two-stage least squares estimator with fixed effects (2SLS-IF) to a Double/Debiased Machine Learning (DML) approach. The DML estimator delivers more accurate out-of-sample predictions of LATEs when treatment effect heterogeneity and selection are driven by observable characteristics, outperforming 2SLS-IF under model misspecification. We also propose an algorithmic procedure for hyperparameter tuning (MLtune) that enhances the stability and generalization of DML predictions. These findings show that flexible machine learning estimators can meaningfully strengthen external validity in IV analyses, though their success depends on the nature of selection across settings.

Keyword

Abstract

We ask what a deep-learning solver should report as the solution of a dynamic model in which agents make both discrete and continuous choices, so that optimal policies jump. One step of policy improvement computed after training recovers the exact jumps from every training objective and network architecture we study, while the raw networks misplace them. Against an exact DC-EGM benchmark, we show how the training objective, the function class, and explicit threshold detection each determine the accuracy of the network itself.

Keyword