Labor Market Policy Assignment at Entry into Unemployment: Analysis Using Machine Learning Techniques and Randomized Controlled Trials
Even though unemployment has been decreasing in recent years in Germany, improving the chances reintegration of unemployed persons into the labor market remains an important challenge for policy makers as well as for the labor market administration. In our proposed project, we want to have a closer look at important tools used by caseworkers at the beginning and during the placement process. Profiling assigns unemployed individuals into a small number of categories. These categories shape the mix of active labor market policy treatments to which the individual is subsequently exposed. We propose to use methods relatively new to economics for a deeper analysis of profiling of the unemployed as well as of heterogeneous treatment effects in experiments closely related to the placement process and thus to profiling issues. During recent years, administrative data sets encompassing unemployed persons have improved in size and quality, and computers have become able to handle larger datasets. Only recently, economists have begun to take an interest in data mining and machine learning techniques to work with big data, of which administrative data are a variant. Among other potentials, big data (and related big data techniques) allow conducting highly specific segmentations and to support human decision making with automated algorithms. A key aim of our proposal is therefore to explore the potential of big data techniques for profiling the unemployed and for the identification of heterogeneous treatment effects: (1) In a first step, we will compare how accurately different econometric and machine-learning techniques predict the duration of unemployment and compare the results with those of a soft profiling conducted by caseworkers. In doing so, we will also address the complex interplay between profiling and the evaluation of active labor market programs. Dependent on the results of the previous analyses, we will aim to start a project to test a tool displaying the predicted duration of unemployed persons within the framework of a randomized field experiment. (2) In a second step, we will use data from a recently conducted unique field experiment that randomly varied the usage of a) integration agreements and b) an additional newly developed placement tool, as well as c) the possibility to receive intensive in-house counselling services of the Federal Employment Agency. These elements of the placement process are tightly related to the soft profiling conducted by caseworkers. An aim of this second step is to identify groups with heterogeneous treatment effects across groups using machine-learning techniques. Our work should not only provide an important scientific contribution to the literature, but will also be of high interest for policy makers and the labor market administration in Germany.
Arne Uhlendorff, PhD (ENSAE ParisTech)