Working Papers
The Potential of Recommender Systems for Directing Job Search: A Large-Scale Experiment (with L. Behaghel, M. Gurgand, S. Dromundo and T. Zuber). Conditionally accepted, Econometrica.
[ Abstract | Draft ]
We analyze the employment effects of directing job seekers’ applications toward establishments likely to recruit. We run a two-sided randomization design involving about 800,000 job seekers and 40,000 establishments, based on an empirical model that recommends each job seeker to firms so as to maximize total potential employment. Our intervention induces a 1% increase in job finding rates for short term contracts. This impact comes from a targeting effect combining (i) a modest increase in job seekers’ applications to the very firms that were recommended to them, and (ii) a high success rate conditional on applying to these firms. Indeed, the success rate of job seekers’ applications varies considerably across firms: the efficiency of applications sent to recommended firms is 2.7 times higher than the efficiency of applications to the average firm. This suggests that there can be substantial gains from better targeting job search, leveraging firm-level heterogeneity.
Improving LATE Estimation in Experiments with Imperfect Compliance (with S. Loewe). 2024.
[ Abstract | Draft ]
The evaluation of many policies of interest (e.g., educational and training programs) inevitably face incomplete treatment group take-up. Estimation of causal effects in these controlled or natural ``experiments with imperfect compliance’’ usually relies on an Instrumental Variable (IV) strategy, which often yields imprecise and thus possibly uninformative inference when compliance rates are low. We tackle this problem by proposing a Test-and-Select estimator that exploits covariate information to restrict estimation to a subpopulation with non-zero compliance. We derive the asymptotic properties of our proposed estimator under standard and weak-IV-like asymptotics, and study its finite sample properties in Monte Carlo simulations. We provide conditions under which it dominates the usual 2SLS estimator in terms of precision. Under an assumption on the degree of treatment effect heterogeneity, our estimator remains first-order unbiased with respect to the Local Average Treatment Effect (LATE) estimand, setting it apart from alternatives in the burgeoning literature on the use of first-stage heterogeneity to improve the precision of IV estimators. This robustness to treatment effect heterogeneity and the potential for precision gains is illustrated using Monte Carlo simulations and two empirical applications. Applying our methodology to the returns to schooling example (where compulsory schooling laws serve as instruments for educational attainment), we document that our methodology reduces standard errors by 12% to 48% depending on specifications.
Who With Whom? Learning Optimal Matching Policies (with T. Kitagawa). 2025.
[ Abstract | Draft ]
There are many economic contexts where the productivity and welfare performance of institutions and policies depend on who matches with whom. Examples include caseworkers and job seekers in job search assistance programs, medical doctors and patients, teachers and students, attorneys and defendants, and tax auditors and taxpayers, among others. Although reallocating individuals through a change in matching policy can be less costly than training personnel or introducing a new program, methods for learning optimal matching policies and their statistical performance are less studied than methods for other policy interventions. This paper develops a method to learn welfare optimal matching policies for two-sided matching problems in which a planner matches individuals based on the rich set of observable characteristics of the two sides. We formulate the learning problem as an empirical optimal transport problem with a match cost function estimated from training data, and propose estimating an optimal matching policy by maximizing the entropy regularized empirical welfare criterion. We derive a welfare regret bound for the estimated policy and characterize its convergence. We apply our proposal to the problem of matching caseworkers and job seekers in a job search assistance program, and assess its welfare performance in a simulation study calibrated with French administrative data.
Occupational Mobility and Retraining: Experimental Evidence on Firms’ Hiring Preferences (with G. Azmat, L. Behaghel, R. Rathelot and J. Sultan). 2025.
[ Abstract | Draft ]
Governments typically address occupational mismatch through two types of interventions: (i) redirection policies, which nudge workers toward tighter labor markets, and (ii) retraining policies, which aim to bridge skill gaps. To assess the effectiveness of these policies in a unique setting, we conducted a large-scale correspondence experiment in France, sending 6,668 fictitious applications across six tight occupations and randomly varying applicants’ training and experience. We find that candidates with both initial training and experience in the target occupation received the highest callback rate, followed closely by movers who completed long retraining programs. Short-retraining and untrained movers received half as many callbacks. We also find that the retraining premium increases with the tightness in the local labor market. These results clarify the relative effectiveness of redirection versus retraining policies; we conclude by discussing the conditions under which the costs of these two policy instruments for the government are offset by savings on unemployment benefits.
Publications
Skill Distance Between Occupations and Post‑Training Professional Transitions of Jobseekers (with D. Mayaux, K. M. Frick and T. Zuber). 2025. Economics and Statistics.
[ Abstract | Draft ]
Does vocational training help correct structural imbalances in the labour market? We propose a new measure of the skills distance between occupations, obtained by fine-tuning a large language model on a sample of job offers. Using this method, we demonstrate that the "return to employment" differential between jobseekers with and without training is driven by a reallocation of workers towards occupations that are very different from their previous posts in terms of the skills required. From a purely reallocative perspective, however, the return to employment differential associated with vocational training does not appear to be driven by more jobseekers moving to occupations where employers are struggling to recruit.
Selected Work in Progress
Exploiting Bounded Treatment Effect Heterogeneity for Improved Inference in (Quasi-)Experiments with Imperfect Compliance (with X. D’Haultfoeuille, P. Ketz and S. Loewe).
[ Abstract ]
As a follow-up research project to Hazard and Loewe (2024), this work consider the same setting while adding a bounded treatment effect heterogeneity assumption. Relying on the constraints imposed by the LATE (Angrist and Imbens, 1994) identifying assumption on the joint distribution of the reduced form and first-stage estimands, we propose a novel estimator based on a projection of empirical moments on the constraint with a high potential for reduction in RMSE. Inference results based on resampling methods---taking into account the bias of the estimator as well as the challenge raised by inference at the border of the parameter space---are currently being developed, with encouraging results in Monte-Carlo simulations and candidate applications.
Evaluating the Effect of Training Programs for the Unemployed: an Examiner Design Approach (with L. Behaghel, M. Gurgand, U. Oyon Lerga).
[ Abstract (early stage project) ]
We exploit the random allocation of caseworkers to job seekers in France---and the heterogeneity in caseworkers' propensity to place individuals in training programs---in order to build an instrument for entering a training program while unemployed. To alleviate threats to the exclusion restriction assumption, we are currently developing an identification approach combining (i) the intuition behind of so-called ``zero-first-stage'' falsification test, (ii) an identification-at-infinity argument and (iii) a single-index assumption imposed on caseworkers' direct impact on individuals' job finding rate (violating the exclusion restriction of the instrument). Our framework lends itself nicely to the use of machine-learning predictions in a first step to identify the zero-first-stage subgroups that are essential for our identification-at-infinity approach.