Kristian U. O. Larsen defends his PhD thesis at the Department of Economics

Candidate:

Kristian Urup Olesen Larsen, Department of Economics, University of Copenhagen

Title:

Gender Inequality, Labor Supply and New Structural Methods 

Supervisors:

Assessment Committee:

Thomas Jørgensen, Associate Professor, Department of Economics, University of Copenhagen
Spencer Bastani, Adjunct Professor, Department of Economics, Uppsala University
Herdis Steingrimsdottir, Associate Professor, Department of Economics, Copenhagen Business School

Summary:

This thesis consists of three self-contained chapters, each made up of a research paper. While the three papers do not share a common topic or method, they are all influenced the CEBI research agenda. Chapters one and two have in common their reliance on the Danish administrative data and the use of micro-econometric methods, while the third chapter carries out methodological work using reinforcement learning. Together the chapters span from dynamic structural modelling through contributions to the estimation of tax elasticities to issues of gender inequality. Each of the chapters can be read independently of each other, as each of them contains all the relevant references and appendices.

Chapter 1: Couples and Gender Inequality

The first chapter documents the transition from single to living in couples as a novel source of gender inequality. When women enter their first cohabiting couple — typically around their early to mid-20s — they experience an average income penalty of 5% of their expected annual income had they continued living as singles. This effect comes on top of an income loss deriving from increased fertility and thus increased exposure to the child penalty. Men are unaffected by the transition, which means couple-formation by itself contribute to the gender inequality in income. I investigate two main theoretical mechanisms with the potential to explain the existence of a cohabitation penalty; household specialization and gender norms. While specialization into homemaker/breadwinner roles does not seem to contribute to the cohabitation penalty, I show evidence that suggests adherence to traditional gender norms, which are transmitted across generations, is driving the magnitude of the penalty.

Chapter 2: Micro vs Macro Labor Supply Elasticities: The Role of Dynamic Returns to Effort

The second chapter, co-authored with Henrik Kleven, Claus Thustrup Kreiner and Jakob Egholt Søgaard, develops a new model of earnings responses to taxes when the returns to effort are dynamic. In this model returns to effort accrue with delay, because performance evaluations in the form of job or occupation switches happen stochastically. We provide descriptive empirical evidence that supports our theoretical model, and some that explicitly contradicts the classical static model of labor supply. In our model, earnings responses around discrete job-switches (partially) identify the long-run macro elasticity of labor supply, while approaches that include job-stayers are attenuated. We provide quasi-experimental evidence on earnings responses to taxes using a Danish tax reform, and exploiting earnings variation due to job-switches. We find that the macro elasticity based on job-switchers is much larger than the micro elasticity, because the standard approach does not account for dynamic compensation effects. This chapter is also available as NBER Working Paper 31549.

Chapter 3: Using Reinforcement Learning for Solving Dynamic Discrete-time Problems in Economics

In the third chapter, co-authored with Joachim Kahr Rasmussen, we show that modern methods from the field of reinforcement learning can be used to solve dynamic problems of high relevance to economists. The traditional approaches to dynamic programming favors precision and exhaustiveness at the cost of being inappropriate for high-dimensional problems. Reinforcement learning on the other hand trades off accuracy for better generalization to higher dimensional problems. This makes RL agents attractive for a certain type of economic problem where dimensionality is a practical constraint and researchers are willing to forego some precision in order to approximate solutions in higher dimensions. We demonstrate applications of Reinforcement Learning across problems with discrete as well as continuous state and action spaces, and find that RL shows promising potential in dealing efficiently with high dimensionalities. We also highlight that time-efficient solutions using RL rely on optimally tuned hyperparameters - a process that in itself can be tremendously costly in terms of computational power. This chapter also appeared in the PhD dissertation of Joachim Kahr Rasmussen.

An electronic copy of the thesis can be requested here: lema@econ.ku.dk