Unemployment is high and rising in many cities in developing countries, particularly among the youth (ILO, 2022). Employment growth in manufacturing exports, a central feature of East and Southeast Asian economic development, has been limited in most African countries (Brandt et al., 2008; Hsieh, 2002; Rodrik, 2018). Migration to high-income countries faces political barriers, limiting another avenue to large-scale employment (Clemens, 2011). Digital services exports offer the prospect of increasing both output and employment in developing economies. Services exports have risen rapidly, roughly doubling between 2005 and 2019 for low-income countries and tripling for middle-income countries (Nayyar & Mulabdic, 2022). In Kenya, where this study takes place, the IT industry has grown by more than 10% a year in recent years (Communications Authority of Kenya, 2019).

Advocates of digital services exports point to many potential advantages for both firms and workers. For firms, these offer export opportunities that are lower-cost and less regulated than goods exports and they do not require large-scale capital investment. If firms can identify, hire, and retain productive workers, digital service exports can be easily scaled. For workers, these jobs offer fewer health and industry risks than manufacturing; potentially flexible hours and locations, which facilitate access for people with family care responsibilities and people living outside large cities; skill development and the potential for career ladders; and easily measurable output, allowing for efficient compensation and promotion systems. However, these claims are almost entirely untested, with the partial exception of call center work. Given the apparent benefits of this sector, it is worthwhile to explore mechanisms to reduce hiring costs to fuel sector growth.

Therefore, the objectives of this project are twofold. Firstly, this research project will estimate the benefits of digital services export jobs in Kenya. To do so, the researchers have partnered with Shortlist, a hiring platform that offers low-cost recruitment services to digital companies. In the context of this study, Shortlist will offer recommendations of randomly selected jobseekers taken from a pool of qualified candidates to their clients. This variation will be exploited to explore the benefits of being employed in the digital sector in the medium term. Secondly, this project will employ machine learning techniques and psychometric assessments leveraging the productivity outcomes of candidates to create an optimal hiring model.

This project takes place in Nairobi, Kenya. Each candidate has to submit an application for “Digital Work Projects” through Shortlist’s application portal. Candidates who go through this application stream will be considered for multiple vacancies across multiple firms within the call center industry. Candidates deemed eligible for jobs based on Shortlist’s application criteria are then randomized into treatment status either being included on a shortlist or not being placed on any list. At the candidate level, this generates random variation in the probability of hiring, which will be leveraged to answer the first research question.

Data will be collected from Shortlist's online application form, Shortlist’s skills assessment, and post-treatment surveys of candidates. These data will be used to determine the impact of being offered a job due to being placed on a shortlist on outcomes like employment, earnings, skill development, and job ladders.

The researchers will also build an optimal hiring model by combining data on candidates' backgrounds and skills with performance data from Shortlist's clients. The randomization process overcomes an important limitation of prior research in this area. Randomization means that candidates with a wide range of characteristics are shortlisted for jobs. Without this randomization, the sample of hired candidates is highly selected on both observed and unobserved characteristics. Models that predict performance in these selected samples may be unreliable for predicting performance for types of candidates not observed in the sample (Madan et al., 2021; Tommasi et al., 2015).