Measuring non-workers’ labor market attachment with machine learning

Studying the labor market attachment (LMA) for the non-working population is crucial for several economic outcomes, such as real wages or long-term non-employment. Official statistics rely on self-reported variables and rule-based procedures to assign the labor market status of an individual. However, this classification does not take into account other individual-level characteristics, like variables related to reservation wages or the amount and type of job offers received, implying that estimates of non-worker status could be biased. In this paper, we propose a novel methodology to measure non-workers’ LMA. Using the Spanish Labor Force Survey (LFS), we define two groups (attached vs. non-attached), and estimate a probability distribution for each individual of belonging to such groups. To recover these probability distributions, we rely on unsupervised and supervised machine learning algorithms. We describe the differences between LFS unemployment, other measures of attachment in the literature, and our non-worker classification. We identify the instances in which our proposed methodology has a tighter relationship with measures like salaries, GDP and employment flows.