One of my file drawer projects that never got off the ground was to conduct an experiment where we trained people in a developing country on how to freelance on global online websites, and see whether this could be a potential substitute to having to migrate to earn higher incomes. Post-Covid and with the increase in remote working there seemed to be a lot more interest in this possibility. Late last year the World Bank released a report Working without borders” which estimated that there are 154-435 million online gig workers round the world, with lots of regional and local gig platforms as well as global ones. I’ve recently come across three papers that look at what the returns are to trying to help more workers access such platforms, which offer differing types of programs and differing results. The studies focus on freelancing (or e-lancing), which tend to be project work that involves intermediate to high-skills such as graphic design, software development, etc. This is distinguished from what Fazio et al. (referenced below) call microwork, which is bite-sized tasks that can be done within seconds or minutes such as image tagging, reading dialog, or data entry. A key question is whether people already have the technical skills needed to compete on such platforms, but need help learning how to get onto such platforms and get contracts, or whether they also need more extensive training on technical skills.
El Salvador (and Haiti): platform training helped get people onto platforms, but had no sustained impacts
Fazio, Freund and Novella have a paper in the JDE (Jan 2025) test a government job training program in El Salvador done with the IADB that tests the effectiveness of providing workers with training and assistance in learning how to engage with these online platforms such as by creating a profile, preparing proposals, interacting with clients etc.
The experiment recruited subjects through online marketing, and required them to have a technical college or university degree or be final year students, or to have at least two years of equivalent experience, with skills in web and mobile development, digital marketing, graphic design, etc. They also had to have at least intermediate level English skills. The sample has an average age of 30 and 55% are women. They have a sample of 711 who met their criteria, and randomly allocated them to:
· A control group of 344 individuals plus a waitlist of 17 who could be used if treated individuals didn’t take-up the program.
· A treatment group of 350 individuals who are offered a 12 week training, of around 5 hours per week online, covering how to navigate and engage with these platforms. This was divided into a theoretical phase, followed by a practical phase to try implementing, as well as some applicants also getting an English for Freelancers course which had industry jargon, CV tips, etc. The program cost $700 per beneficiary, with $550 of this dedicated to tutors to help students.
They measure outcomes through administrative data on the treatment group, and through a one-year phone follow-up survey that had an attrition rate of 24.5% (10 p.p. higher for control than treatment).
They find the following:
1. It is hard to get people to take-up and follow-through on online training: despite people self-selecting into the course, and them assigning them tutors to help encourage and follow-up, they find only 39% completed the theoretical phase, 16% completed the full program, and while 45% started the English course, only 16% completed it.
2. Training does get more people onto the platforms and getting a first contract: 51% of the control have an online freelancing profile, and 10.7% have signed at least one job contract in the past year online. The ITT is for a 27.2 percentage point (p.p) increase in having a profile, and 5.8 p.p. increase in getting a job contract – which are large relative to the control means.
3. Most people don’t end up getting many contracts, and there is no sustained success on the number of contracts or labor income: On average participants only received 1.4 online job offers in the entire year, and signed 0.8 contracts, and the share of their income from freelancing was less than 4 percent. Treatment has no significant impact on how many contracts they get, how much income, or whether they expect to keep freelancing in the next year.
The authors suggest that one reason for a lack of sustained impact could be poor ratings. They note that both treatment and control only average a rating of around 3 out of 5, whereas the average freelancer rating for all gig-workers on Upwork is 4.9. A similar lack of impact is seen in a pilot of the same program done in Haiti with 94 workers, published in Economic Letters.
Bangladesh: training women in technical skills did deliver benefits from a very low base
A recent working paper by Das et al. tests the impact of training women in the technical skills needed to be able to provide services on these platforms: focusing on graphics design, digital marketing, and web research.
They recruited women aged 18 to 35 in Bangladesh online and through circulars in schools and universities. Women had to have a minimum of secondary school certification, and then pass a literacy test, and computer test. Average age is around 26. Only 19% were engaged in any earnings activity at baseline. They have a sample of 900 who are randomly allocated to:
· A control group of 450 women
· A treatment group of 450 women who are offered a 4 month training program that takes place in-person at first, then online due to Covid, three times a week for 4 hours at a time, for a total of 50 classes. Participants could choose which one of graphic design, digital marketing, or web research and support they wanted. Sessions included both theoretical and practical training, and then afterwards there were two months of mentoring and guidance on how to develop their profiles on Fiverr and Upwork. The cost is 35,000 BDT (around $423 at the time).
They have a 13 month follow-up survey that was conducted through a mix of in-person and phone surveys, which had a response rate of around 66% (although some attrition happened at baseline too, and so it is 28% attrition relative to baseline, with a 3 p.p. difference between treatment and control).
They find:
1. Higher take-up both in-person and online: 76% started the training, 72% completed at least half, and 34% completed it all. The average number of classes for those who participated online was actually higher than in-person (31 vs 22), although this could also reflect changes in the opportunity cost of time during Covid.
2. Lasting impacts on the likelihood of working, and on pay: 53% of the control group are employed at follow-up, and the treatment impact is 9.9 p.p; this increase comes from freelancing work, which more than doubles, with a treatment effect of 15.8 p.p. compared to 11.9% doing it in the control group. Monthly income earned from freelancing increases from 779 BDT in the control group by an additional 886 BDT in the treatment group. That is, they seem to be earning about $10-11 more per month. The authors explore whether this increase in income is coming from just an extensive margin effect of getting more people online, or also through improving productivity once they are online. The bounds are wide, but under certain assumptions about how the incomes of “always workers” compare to compliers, there seems to be some productivity effect.
3. These benefits flow through into household outcomes: spending on food is higher, and is spending on mobile and internet (needed for work, but with likely other beneficial impacts for household members). Impacts appear higher for those who are current students, with no substitution away from their existing schooling.
The main mechanism here seems to be through improved skills – treated individuals self-report having higher technical skills in areas like word processing, spreadsheet, web design, etc. Unfortunately we do not see any information on what they are doing on the platforms, so we don’t see how many projects they are doing, what types of employers they are working for, what their ratings are, and how skewed the distribution is. While the $10 per month increase is large in relative terms, it is still small relative to the $423 cost of training, and impacts would have to persist well beyond the time horizon of this study to make it worth doing – and do not seem large enough to enable a worker to borrow to pay for the training and earn back the cost in a couple of years, especially once interest rates are included.
[postscript: Narayan Das, one of the authors, noted that the gain in income for those who take-up the training is more like $25 per month once you look at the TOT and also include a bit of extra earnings in non-freelance work - but she notes is still not high enough to enable trainees to borrow against these future earnings to finance training]
Reflections
Taken together these studies show that it is perhaps not enough to help workers get onto these platforms, but you might need to help them bring their technical skills up to a level to be globally competitive. With recent developments in AI, it is also unclear how many of these graphic design and web design jobs will still exist. Upwork reports that generative AI is lowering the value and number of less skilled tasks, but providing more opportunities for more complex, high-value contracts. The bar to train people in developing countries without these skills to the level needed to compete may therefore now be higher. Given that both the above studies already had relatively small and specialized samples, it seems very challenging to design such programs to provide new job opportunities to tens of thousands of job-seekers in a country. But yet many workers are using such platforms even without government programs to help them.
Finally, seeing these experiments also makes me more skeptical about the likelihood that these types of programs could be used as a substitute to stem the flow of irregular migration. It is very hard to identify in advance who is at the highest risks of irregularly migrating, and then also getting enough of these same people to attend training and be successful enough on an online platform to have sustained income is multiplying a lot of small probabilities together.
Join the Conversation