IAB-Colloquium (en) Archive - Page 9 of 17 - IAB - Institut für Arbeitsmarkt- und Berufsforschung

How can AI improve search and matching? Evidence from 59 million personalized job recommendations

This lecture is about how artificial intelligence can be used to reduce friction in markets.

We explore how Artificial Intelligence can be leveraged to help frictional markets to clear. We design a collaborative-filtering machine-learning job recommender system that uses job seekers' click history to generate relevant personalised job recommendations. We deploy it at scale on the largest online job board in Sweden, and design a clustered two-sided randomised experiment to evaluate its impact on job search and labour-market outcomes. Combining platform data with unemployment and employment registers, we find that treated job seekers are more likely to click and apply to recommended jobs, and have 0.7 percent higher employment within the 6 months following first exposure to recommendations. At the job-worker pair level, we document that recommending a vacancy to a job seeker increases the probability to work at this workplace by 10 percent. We propose a decomposition exercise of the net employment effects into three channels. The most important channel corresponds to the increase in the number of applications due to recommendations (first channel), partly offset by the lower conversion into employment of marginal applications (second channel). Congestion effects (third channel) are not a significant contributor to the overall effect. We also find larger employment effects when recommended vacancies are less popular, and for recommendations that broaden search further away in geographical and occupational distance.

The Direct and Indirect Effects of Online Job Search Advice

We study how online job search advice affects the job search strategies and labor market outcomes of unemployed workers.

We study how online job search advice affects the job search strategies and labor market outcomes of unemployed workers. In a large-scale field experiment, we provide job seekers with vacancy information and occupational recommendations on an online dashboard. A two-stage randomized design with regionally varying treatment intensities allows us to account for treatment spillovers. Our results show that online advice is highly effective when the share of treated workers is relatively low: in regions where less than less than 50% of job seekers are exposed to treatment, working hours and earnings of treated job seekers increase significantly in the year after the intervention. At the same time, we find substantial negative spillovers on other treated job seekers for higher treatment intensities, resulting from increased competition between treated job seekers who apply for similar vacancies.

Designing labor market recommender systems: the importance of job seeker preferences and competition

This paper questions the design of job recommender systems (RS).

This paper questions the design of job recommender systems (RS). We argue that state of the art ML-based algorithmic recommendations aimed at identifying a hiring score from past successful matches do not always result in improved outcomes for job seekers (JS). This is because first, the objectives of these recommendations do not align with the ones of the JS and second, they are usually generated independently of each other, without considering competition. Using a theoretical model of a two-sided market with an application stage, we discuss the needs that RS should meet. We then show that the ML-based hiring score, from which recommendations are typically derived, is only one of the necessary ingredients to meet these needs. Additionally, a matching score between JS and job offer profiles must be considered, and the two should be combined to form a criterion that reflects the expected utility. Our empirical analysis confirms this quantitatively matters, using the RS designed as part of a long-term project in collaboration with the French Public Employment Service. This project leverages extensive and detailed data on applicants, firms, and past job searches. Moreover, we discuss how optimal transport can be leveraged to design RS that avoid congestion, viewing the recommendations as a collective problem rather than a series of individual programs.

Labor Demand and Workforce Diversity: Evidence from Two Natural Experiments

We address the question whether diversity in the workforce increases in the absence of diversity-targeted policy interventions when academic labor markets become tighter.

Increasing diversity is one of the challenges in modern labor markets. Increased representation, participation and inclusion in the workplace may not only desirable from a societal or political, but also from an economic perspective. The lack of diversity has been documented to be particularly salient in the academic sector. In this paper, we address the question whether diversity in the workforce increases in the absence of diversity-targeted policy interventions when academic labor markets become tighter. We focus on the German academic labor market for professors which has been characterized as a slack labor market with an over-representation of males in which closed networks play an important role. To study market forces, we explore two natural experiments that unexpectedly increased labor demand and led to tighter labor markets for professors. First, using newly digitized data from the German Federal Statistical Office on academic staff during the German university expansion in the 1960s and 70s, we document an increase in the share of female professors from 0.62 percent in 1960 to 4.39 percent in 1977. Second, we explore between-discipline variation in staff replacements at universities in East Germany after the reunification. Using administrative data on university staff, we find that nine years after the fall of the German wall professors are significantly younger in the Social Sciences (strongly affected by replacements) compared to STEM subjects (barely affected), in the East relative to the West. There is no respective significant change in the share of female professors. However, professors have a more diverse academic background as measured by their university of habilitation. Taken together, our analyses demonstrate that positive labor demand shocks indeed have the ability to contribute to more diversity in academia in some dimensions and, by market force and in the absence of targeted policy interventions, break up some of the "Old-Boys' Club’'.

Estimating returns to special education: combining machine learning and text analysis to address confounding

Results from the first ever study to examine short- and long-term returns to special education programs with causal machine learning and computational text analysis methods.

Leveraging unique insights into the special education placement process through written individual psychological records, I present results from the first ever study to examine short- and long-term returns to special education programs with causal machine learning and computational text analysis methods. I find that special education programs in inclusive settings have positive returns in terms of academic performance as well as labor-market integration. Moreover, I uncover a positive effect of inclusive special education programs in comparison to segregated programs. This effect is heterogenous: segregation has least negative effects for students with emotional or behavioral problems, and for nonnative students with special needs. Finally, I deliver optimal program placement rules that would maximize aggregated school performance and labor market integration for students with special needs at lower program costs. These placement rules would reallocate most students with special needs from segregation to inclusion.

The Care-Dependent are Less Averse to Care Robots: An Empirical Comparison of Attitudes

Using a large-scale experimental vignette study, we investigated people’s affective attitudes toward care robots.

A growing gap is emerging between the supply of and demand for professional caregivers, not least because of the ever-increasing average age of the world’s population. One strategy to address this growing gap in many regions is the use of care robots. Although there have been numerous ethical debates about the use of robots in nursing and elderly care, an important question remains unexamined: how do the potential recipients of such care perceive situations with care robots compared to situations with human caregivers? Using a large-scale experimental vignette study, we investigated people’s affective attitudes toward care robots. Specifically, we studied the influence of the caregiver’s nature on participants’ perceived comfort levels when confronted with different care scenarios in nursing homes. Our results show that the care-robot-related views of actual care recipients (i.e., people who are already affected by care dependency) differ substantially from the views of people who are not affected by care dependency. Those who do not (yet) rely on care placed care robots’ value far below that of human caregivers, especially in a service-oriented care scenario. This devaluation was not found among care recipients, whose perceived level of comfort was not influenced by the caregiver’s nature. These findings also proved robust when controlled for people’s gender, age, and general attitudes toward robots.

Performance costs and benefits of collective turnover: A theory-driven measurement framework and applications

We model the effect of collective turnover on workplace performance as the sum of its costs and possible benefits occurring through changes in workforce match quality.

Building on job matching theory, we model the effect of collective turnover on workplace performance as the sum of its costs and possible benefits occurring through changes in workforce match quality. The resulting theoretical turnover-performance relationship is generally curvilinear, nesting all the hitherto known patterns -- linear, ``U-shape'' and ``inverted U-shape'' -- as special cases. We show how one can estimate this relationship empirically, for matched worker-plant data, and calculate the implied costs and benefits of turnover. Applications to data from two retail networks reveal that turnover is more costly than beneficial.

Automated classification for open-ended questions + Hammock Plots

We investigate whether BERT is more effective at automated coding of answers to open-ended questions than non-pre-trained statistical learning approaches.

Answers to open-ended questions are often manually coded into different categories. This is time consuming. Automated coding uses statistical/machine learning to train on a small subset of manually coded text answers. The state of the art in NLP (natural language processing) has shifted: A general language model is first pre-trained on vast amounts of unrelated data, and then this model is adapted to a specific application data set. After reviewing some earlier results, we empirically investigate whether BERT, the currently dominant pre-trained language model, is more effective at automated coding of answers to open-ended questions than non-pre-trained statistical learning approaches. In the second part of the talk, I discuss the hammock plot for visualizing categorical or mixed categorical data.

Household Labor Supply Elasticities: Evidence from Cross-Border Workers

We provide evidence for intensive margin adjustments in hours worked consistent with estimates.

After the Swiss National Bank unexpectedly abandoned a minimum exchange rate policy in 2015, the Swiss franc appreciated by more than 10 percent against the Euro. The appreciation implied a sudden increase in real wage incomes for over 40,000 German cross-border commuters into Switzerland. We use this exchange rate shock to estimate the own-wage and cross-spouse labor supply elasticities from administrative tax returns data and find a 5% drop in taxable income for cross-border workers and a 1.5% reduction in taxable income for cross-border worker spouses. We provide evidence for intensive margin adjustments in hours worked consistent with these estimates.

Longer careers: A barrier to hiring and coworker advancement?

We study a Dutch reform that raised the retirement age by 13 months and nearly tripled employment at targeted ages.

Government policies are encouraging older workers to delay retirement, which may curb younger workers' career advancement. We study a Dutch reform that raised the retirement age by 13 months and nearly tripled employment at targeted ages. Using monthly linked employer-employee data, we show that affected firms delay and decrease replacement hiring, and coworkers' earnings fall via reductions in hours worked, wages, and promotions. The hiring and coworker spillovers offset most of the additional hours worked by older workers. These spillovers exacerbate within-firm earnings disparities, redistributing earnings from low to high earners, young to old workers, and women to men.