Equalizing the Effects of Automation? The Role of Task Overlap for Job Finding
Joint with Sabrina Genz and Emilie Rademakers.
This paper investigates whether task overlap can equalize the effects of automation for unemployed job seekers displaced from routine jobs. Using a language model, we establish a novel job-to-job task similarity measure. Exploiting the resulting job network to define job markets flexibly, we find that only the most similar jobs affect job finding. Since automation-exposed jobs overlap with other highly exposed jobs, task-based reallocation provides little relief for affected job seekers. We show that this is not true for more recent software exposure, for which task overlap mitigates the distributional consequences.
Fissured Firms and Worker Outcomes
Joint with Matias Cortes, Ana Oliveira and Anna Salomons
We consider how firms’ organization of production relates to workers’ wages. Using matched employer-employee data from Portugal, we document that, within detailed industries, firms differ starkly in their occupational employment concentration, with some firms employing workers across a broad range of occupations and others being much more specialized. These differences are robustly predictive of wages: a worker employed in a specialized, i.e. ‘fissured’ firm earns less than that same worker employed in a less specialized firm; and being employed in a specialized firm predicts significantly lower subsequent wage growth, driven by reduced hourly wages as well as lower hours worked by workers remaining in specialized firms. This wage penalty is observed across a wide range of occupations. Firm specialization helps account for the role of firms in inequality: over half of the wage penalty from specialization is explained by differences in firm productivity, and firm specialization is strongly negatively related to AKM firm fixed effects. Lower rent-sharing also contributes to the observed wage penalty from firm specialization.
Automatically classifying job titles into occupations using FastText
and socio-demographic information
I test the effectiveness of FastText, a word embedding algorithm, to classify job titles
into a standarized list of occupations. Furthermore, I study the incorporation of
individual’s information to enhance the classification results. To do so, I develop a
deep learning classification algorithm that combines pretrained word embeddings with
other numerical variables. I find that in its best configuration FastText achieves a
macro F1 score of 0.73, while the deep learning classification algorithm that uses individual
characteristics improves this score to 0.85 when classifying into 222 detailed
occupations.