Online Platforms and the Labour Market: Learning (with Machines) from an Experiment in France - DUMAS - Dépôt Universitaire de Mémoires Après Soutenance Access content directly
Master Thesis Year : 2019

Online Platforms and the Labour Market: Learning (with Machines) from an Experiment in France


I study the effect of an online job search assistance program taking advantage of a previousexperiment made by the French public employment services, which provides some exogenousvariation in the use of this platform. I focus on the heterogeneity analysis of this treatment,using two main different approaches.The first one is theory-driven, and focus on the analysis of the heterogeneity of thetreatment with respect to various different labour market tightness indicators. Two mainassessments can be made based on this analysis. (i) Tightness indicators are (surprisingly)decorrelated, making it difficult to corroborate the rare significant results obtained. (ii) Theset of significant results obtained suggest that the treatment effect isincreasingin labourmarket tightness. I suggest competing ways of modelling the treatment consistent withthose results. I also document some evidence of a larger treatment effect for individuals withweaker employment prospects. This is in line with other empirical evidence in the literatureevaluating job search assistance programs.The second approach is more data-driven, and resorts to the new machine learning (ML)techniques developed for heterogeneity analysis. I focus on tree-based techniques and forests,which have been central in the development of these techniques. The results of this analysisshed light on the limits of ML in the exploration of treatment effect heterogeneity, especiallyas the main ML-specific test for treatment effect heterogeneity developed by Chernozhukovet al. (2018a) concludes that ML is unable to detect any heterogeneity — yet this mightbe not that surprising after all given the lack of statistical power (low take-up) and theprobably low order of magnitude of the treatment effect studied. Still, I provide applicationsof a large part of the existing ML techniques for treatment effect heterogeneity, trying totake advantage of each of them to document which are the dimensions that are likely to beimportant to study treatment effect heterogeneity in my setting.
Fichier principal
Vignette du fichier
MT_201903_.pdf (2.19 Mo) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

dumas-02407555 , version 1 (12-12-2019)


  • HAL Id : dumas-02407555 , version 1


Yagan Hazard. Online Platforms and the Labour Market: Learning (with Machines) from an Experiment in France. Economics and Finance. 2019. ⟨dumas-02407555⟩
258 View
159 Download


Gmail Facebook X LinkedIn More