As markets increasingly (and often exclusively) take place online, they generate both new types of data as well as new challenging research questions. The Internet as a data source for research in labor economics is therefore a focal point of IZA’s research data center, IDSC. Organized by Nikos Askitas and Peter J. Kuhn, the two-day workshop showcases innovative multidisciplinary research with data from internet job boards, one of the main modes of matching facilitation in labor markets worldwide today.
On the heels of both technological hype and techno-optimism, as well as aiming at more tangible benefits such as cost reduction and speed, a market emerges in which firms offer algorithmic matching services to hiring firms. In the workshop’s first keynote, Manish Raghavan presented the current state of affairs in this algorithm-driven job matching market with a focus on compliance with anti-discrimination law in the US.
Firms offering algorithmic hiring services have to consider and filter both on the quality of the training data used by the algorithms as well as the prediction targets and evaluate risks and trade-offs, a new and complex problem. Among the risks involved is the emergence of a monoculture: While in the pre-algorithmic hiring world there is a variation of (possibly error-prone) manual hiring procedures, in an algorithmic universe all hiring firms using the same algorithms commit the same errors – to the potential detriment of the labor market and society at large.
Recommender systems in two-sided markets
In the workshop’s second keynote, Thorsten Joachims, who also works on the border of computer science and economics, discussed research designed to produce fair rankings from biased data in two-sided markets. Search engines and recommender systems have become the dominant mode of matchmaking in a wide range of two-sided markets, such as retail, entertainment, employment, or even romantic partners. Consequently, such systems can shape markets. Distortion of opportunity allocation to market participants can occur either due to exogenous reasons, such as biased training data, but also due to reasons endogenous to the machine learning algorithms. Removing such distortions is therefore a new and important problem.
Defying distance? Provision of services in the digital age
In her paper, Amanda Dahlstrand-Rudin studies how digital platforms are transforming service provision in health care by making the physical distance between provider and user less relevant. Using data on 200,000 patients and 150 doctors, she first analyzes the effect of the random assignment of patients to primary care doctors that took place when Sweden moved these services online. Random assignment improved aggregate health outcomes, in part because it increased the exposure of high-risk patients to doctors who were better able to treat them.
Dahlstrand-Rudin then goes further, using the estimated causal effects derived from random assignment to project the possible health care benefits of using existing online information to actively match patients at high risk of avoidable hospitalizations to doctors skilled at triaging. This would reduce avoidable hospitalizations by an additional 20 percent. Overall, the study dramatically shows how moving service provision to online platforms has the potential to improve service quality while reducing inequality at the same time.
RCTs on job seekers
The study presented by Jung Ho Choi measures how information about the diversity of a potential employer’s workforce affects individuals’ job-seeking behavior, and whether workers’ preferences explain corporate disclosure decisions. By embedding a field experiment into job recommendation e-mails sent from a leading U.S. career advice agency, the authors find that disclosing company diversity scores in job postings increases the click-through rate and willingness-to-pay of job seekers for firms with higher diversity scores.
Using a follow-up survey, the researchers also demonstrate that diversity information is more valuable to female job-seekers and people of color. The results provide useful new insights into how U.S. firms are likely to respond to growing pressure for firms to voluntarily disclose diversity metrics in their 10-K reports under new SEC disclosure requirements.
See the workshop program for more information.