I develop a framework for modeling the lifetime dynamics of an individual's health and health care use that can be employed to estimate the aggregate, distributional, and welfare outcomes of related policies and trends.  The structural methodology innovates upon previous approaches by integrating the life-cycle evolution of an objective measure of health with the dynamics of health care use identified along both necessary-discretionary and chronic-acute dimensions while capturing heterogeneity across age.  These processes are embedded in a life-cycle model, which I use to quantify the impacts of projected demographic trends and public health insurance reforms on the U.S. economy while highlighting the motivation and consequences of the model's innovations to rigorously estimate the linked dynamics of health and health care use.  Population aging increases necessary health care use resulting in a 20 percent increase in total health care expenditure as a percentage of GDP by the year 2070.  Expanding Medicare's generosity increases discretionary health care use, decreases household saving, and increases elderly welfare via reducing the financial risk of necessary health care.

This paper presents a tractable method for predictive modeling with high-dimensional, spatio-temporally-distributed economic data via borrowing techniques from visual recognition machine learning.  Specifically, I cast the time series of spatially disaggregated U.S. economic data as a temporal sequence of geographic `images' in a computer vision setting and develop a deep learning model architecture to evaluate whether leveraging the spatio-temporal distribution of data features can improve macroeconomic forecasts.  The resulting spatial recurrent convolutional neural network model accurately forecasts changes in U.S. GDP over a long time horizon both for in-sample and out-of-sample data and outperforms more traditional linear methods as well as deep learning models that do not utilize the data's spatial distribution.  Analysis of the estimated model provides intuition for the model's improved performance by highlighting its ability to focus on regional economic experiences and shift its geographic focus over time.

This study applies state-of-the-art machine learning (ML) techniques to forecast IMF program requests by countries, analyzes the ML prediction results relative to traditional econometric approaches, explores non-linear relationships among predictors indicative of IMF program requests, and evaluates model robustness with regard to different feature sets and time periods.  ML models consistently outperform traditional methods in out-of-sample prediction of new IMF arrangements with key predictors that align well with the literature and show consensus across different algorithms.  The analysis underscores the importance of incorporating a variety of external, fiscal, real, and financial features as well as institutional factors like membership in regional financial arrangements.  The findings also highlight the varying influence of data processing choices such as feature selection, sampling techniques, and missing value imputation on the performance of different ML models and therefore indicate the usefulness of a flexible, algorithm-specific approach.  Additionally, the results reveal that models that are most effective in near and medium-term predictions may tend to underperform over the long term, thus illustrating the need for regular updates or more stable – albeit potentially near-term suboptimal – models when frequent updates are impractical.

Given the uniquely-high private financial costs of physician training and uniquely-low amount of practicing physicians in the United States, this paper constructs a life-cycle model with a health care service sector, physicians, and health uncertainty in order to estimate the welfare implications of publicly-funding the financial costs of physician training.  In the model, publicly-financing medical school tuition increases the number of U.S. physicians per capita to around the OECD average and can alone account for over 12.5 percent of the difference between U.S. health care service expenditure as a percentage of GDP and that of France and Germany.  Direct cash transfers to medical students generate the largest decrease in the price of health care services (11.7 percent) due to the corresponding increase in physician labor.  Because a household’s health care use is uncertain, the decrease in the price of health care services reduces household exposure to financial risk and produces welfare gains - a form of implicit insurance.

Expenditure on non-clinical, administrative tasks in the U.S. health care system is estimated to constitute 4.6 percent of GDP - almost 60 percent of which is specifically billing and insurance-related (BIR).  I construct a life-cycle model with uncertain health care utilization and a health care sector employing both clinical and administrative inputs in order to estimate the macroeconomic consequences and welfare implications of administrative complexity in U.S. health care.  I quantify these impacts via comparison to a counterfactual U.S. economy that has the BIR administrative characteristics of Canada’s health care system.  With Canadian BIR complexity, the price of health care services is 15.8 percent lower and health care service expenditure as a percentage of GDP decreases by 2.1 percentage points.  In the model, differences in BIR administrative complexity can therefore account for over 40 percent of the difference between health care service expenditure as a percentage of GDP in the U.S. compared to Canada.

Macroeconomic data at the country level is restricted by small sample sizes due to infrequent time intervals of observation and missing data – most notably for lower-income and developing countries.  These characteristics complicate efforts to predict related country outcomes, especially when applied to outcomes that are rare such as the occurrence of economic crises.  I examine the potential for harnessing generative machine learning to address these three issues – small sample sizes, missing data, and class imbalance due to rare outcomes – via generative data augmentation.  Specifically, I evaluate whether generative machine learning advancements such as generative adversarial networks (GANs) and variational autoencoders (VAEs) can improve upon the predictive performance of more standard imputation and resampling methods when forecasting the onset of currency, fiscal, and sovereign debt crises within countries.