Are Soft Skills Enough? Experimental Evidence on Skill Complementarity for College Graduates

Abstract

We study how complementarities in skill may affect the returns to vocational training using a randomized controlled trial in Cairo, Egypt. Participants, who were college-educated, were either given a 4-week training in soft skills (e.g., grooming, time management), technical skills (e.g., Microsoft programs, English language), or a mix of the two (half of each). We find large differences in outcomes between the three treatments. The technical and mixed treatments do best in the short term, raising first-job income by about 15%, relative to the soft-skill treatment. In the longer term, the mixed-skill treatment significantly outperforms the other two treatments, giving participants 20-27% higher income. The high returns for this group may come from increased access to jobs that require speaking English, which may be at higher-quality employers. Overall, the results suggest that curriculum details play an important role in the outcomes of vocational training programs and that leveraging skill complementarity can yield tangible benefits.

Introduction

Youth unemployment is a major challenge, especially in developing countries where jobseekers may lack a strong formal education. Training programs are seen as an important step to gainful employment. However, these programs often have only modest returns(Card et al. (2010); Card et al. (2018)). Some have suggested that this may be due to their focus on narrow sets of skills(Krueger and Kumar (2004); Hanushek et al. (2017)). In response, recent work has studied whether teaching “soft” skills, rather than just technical skills, can improve outcomes in training programs. These soft skills are often cited by businesses as being important and difficult to find in jobseekers, and prior research shows that their importance for workers is growing (Deming (2017)).

In this paper, we use a randomized controlled trial (RCT) in Cairo, Egypt, to study how complementarities in skills may affect the returns to vocational training.^[1] Specifically, we compare the effectiveness of 4-week (120-hour) training programs in mostly “soft” skills (e.g., grooming, time management, and listening skills), mostly “technical” skills (e.g., Microsoft programs, technical English language), and a mix of the two (about half of each). The mixed treatment is inspired by the idea that soft and technical skills may be complementary (Weinberger (2014); Cunha et al. (2006)). We test these ideas in a setting where soft skills, rather than technical skills, are likely to be the binding constraint; unusually for a training intervention, the participants in our study were well-educated, having all graduated from college.

It is unclear ex-ante how important skill complementarity may be in practice. The danger with training focused on only one type of skill is the opportunity cost. It is possible that spending most of one’s time learning one particular type of skill can have positive effects, but it could also have negligible or even negative effects if some of that time should have been spent learning other valuable skills instead.

In our context, the soft skills training may improve outcomes for trainees by improving their time management skills, but it is possible that too much focus on these issues may lead to preparation paralysis that would actually decrease productivity. Similarly, the technical skills provided in the training could lead to improvements in productivity through better computer skills and improved technical English skills, which are often prized in developing economies. On the other hand, spending too much time trying to computerize a task or master proper English grammar could be unnecessarily inefficient. Combining both technical and soft skills into one training of equal length may allow workers to maximize the benefits of two types of skills. On the other hand it might lead to them missing out on the most important aspects of either skill since they will only have half as much time to spend on each.

We find that the types of skills taught in vocational training programs matter for outcomes, and the results show the importance of complementarity between soft and technical skills. The longer-term benefits of training are largest for the group that receives a mix of soft and technical skills training. There is no single “return to vocational training”, but rather the returns depend critically on what is being taught and in what proportions.

Our paper makes four main contributions. First, our experiment includes a mixed treatment of soft and technical skills, allowing us to look at complementarities between the two types of skills. In our setup, we can compare a program with roughly equal amounts of soft and technical training to those with almost all soft skills and almost all technical skills. Recent studies on soft skills training compare a treatment either to a control group (e.g., Adhvaryu et al. (2021), Groh et al. (2016)) or to a control group and a technical skills treatment (e.g., Barrera-Osorio et al. (2023)).

Second, our sample is made up of four-year college graduates. This is unusual for a vocational training program because well-educated individuals are typically in less need of training to find employment. In theory, this group should already have high levels of technical skills, making it more likely that soft skills training would have the highest returns. Instead, the soft skills training is outperformed by the other two treatments. However, soft skills continue to be an important part of success for our sample when combined with technical skills for the mixed-skill treatment group. We expect that finding ways to combine the right set of technical and soft skills will be important for designing effective training programs for job-seekers of different skill levels (e.g., those with less than a college education).

Third, we contribute to the literature related to the benefits of English language skills in developing countries. Previous work has suggested that knowing English has benefits in these settings, but evidence from randomized trials on this question is rare or nonexistent. Our results suggest that access to jobs that require English can be an important part of the return to vocational training.

Fourth, in addition to labor market outcomes, we measure the impact of training on mental health, attitudes, expectations, and behaviors. We find some evidence that training can alter people’s expectations and behavior -- particularly marriage plans and job search activity -- even if the labor market effects of the program are small. These findings encourage more longer-term study of the impacts of training.

Study Context, Experimental Design, and Sample Characteristics

Our study takes place in the greater Cairo area of Egypt, a middle-income country with a PPP-adjusted GDP per-capita of about $12,000. Cairo’s population is well-educated, with about a third of individuals aged between 21-35 having at least college degree (Assaad and Krafft (2013)), at least partially due to the fact that public universities are free (Elbadawy (2015)). Despite this, in 2016, the year our first cohort was trained, Egypt faced a 33.4% youth unemployment rate, among the highest of any country (\firstsecondILO (2016)).

There is reason to believe that training in both soft and technical skills could have positive labor market impacts in Egypt. As reported in Osman et al. (2022), in a survey of about 1,000 establishments, 40% of employers believe that there are not enough entry level job applicants with the right skills. When asked which kinds of skills are hardest to find, there is an almost even split, with 46% of firms saying that soft skills are the hardest to find and 54% saying technical skills are harder to find.

We worked with an NGO called Education for Employment Egypt (EFE-Egypt, or EFE). EFE-Egypt was founded in 2007 and has been running training programs in the country since that time. Prior to working with us, they already had extensive experience in both soft and technical skill training. Their focus is on producing high quality “demand-driven” trainings to unemployed and underemployed young adults, working closely with employers to design curricula that match the needs of the market, focusing exclusively on high-quality formal job opportunities for college graduates. The trainings are primarily funded by development organizations (e.g. USAID, UNDP, IDRC, etc.).

Experimental Design

We worked with EFE to design and test the relative effectiveness of training programs with different curricula. Each of the three treatment arms was a 4-week, 120-hour training program made up of a combination of different training modules that were regularly offered by EFE in their programs. Some of these modules focused on “technical skills”, including Microsoft programs (i.e., Word, Excel, and PowerPoint) as well as English vocabulary, grammar, and pronunciation. Other modules focused on “soft skills”, including business skills (i.e., grooming, time management, and interpersonal skills), customer service, email etiquette, and body language.

While EFE insisted on including some soft and some technical training in every treatment arm, we constructed three different treatments with very different mixes of the two types of skills, as outlined in 6. The soft skills treatment consisted of 120 hours of training, 84% of which was from the soft skills modules. The technical skills treatment drew 93% of its 120 hours from technical modules. The mixed treatment, also 120 total hours, was 55% technical and 45% soft. Because the three treatment arms were the same total number of hours, we can isolate the effects of different mixes of skills in our analysis.^[2]

To classify the various EFE modules as either soft of technical, we generally followed EFE’s own classification for each module, although we made the final decisions.^[3] Soft skills, as defined by Heckman and Kautz (2012), are “personality traits, goals, motivations, and preferences that are valued in the labor market, in school, and in many other domains”. This includes skills like time management, interpersonal interactions, teamwork, and conflict resolution. They are considered important in contexts like business and customer service (Schulz (2008); Dixon et al. (2010)), and their importance for workers has grown over time (Deming (2017)).^[4] Technical skills, on the other hand, refer to knowledge and abilities needed to perform specific tasks. These are more likely to be taught in formal education, such as math, writing, language, and computer skills.

The English language curriculum was split into “business English” and “pronunciation” modules. Business English focused on business-specific vocabulary, the various tenses of English, grammar rules, and how to construct correct sentences and questions that might be useful in a business or customer service context. The pronunciation module focused on the sounds of certain letters and letter combinations in English. As this is a foreign language, these modules are clearly “technical”. Similarly, the lessons on Microsoft programs are technical. They are distinct from lessons about grooming, socializing, and how to work in teams, which are counted as soft skills.

6 shows the differences in content between the treatment arms. The soft and technical skill arms are quite different, with the soft skills training including business skills (e.g., dressing, grooming, ice-breaking activities, listening skills, communication styles, interpersonal skills, teamwork activities, and practice in giving presentations), career directions (interview and resume basics), and customer service (how to engage with and influence a customer), whereas the technical arm focuses mostly on English vocabulary and technical skills specific to the retail sector. The mixed-skills arm gets rid of the customer service module and keeps reduced versions of the other soft and technical modules. The partial overlap in content between the treatment arms, such as all three including English pronunciation and Microsoft programs, will make it more difficult for us to find significant differences in outcomes between the three treatments. Appendix Tables 2 and 3 show the various learning sessions included in each module.

To recruit the trainees, EFE advertised through various channels and through their alumni network. Interested individuals would come to their office to apply and fill out an intake form. Following their usual policies, EFE would screen out individuals who had not completed a university degree, and those who were perceived to be “well-off” (i.e., those that went to private universities, or showcased other forms of wealth such as having their own private car).^[5] After EFE recruited enough applicants to fill up three distinct training classes (and a control group), they would send the research team a list of all the eligible trainees and we would randomly split them into four groups: soft skills training, technical skills training, mixed training, or control (no training).

There were a total of 11 cohorts, and hence 11 randomizations. Two cohorts were small, and so they were randomized into only two groups, either soft skills or technical skills. There were six cohorts where the number of applicants was not perfectly divisible by four, and we would allocate the uneven people into training instead of control because the partner was keen to hit their targets on the number of trainings implemented. This explains why those groups are larger than the control group and the combined skills group. The first cohort was trained in November 2016, while the final cohort was trained in October 2018.^[6]

Data Collection and Sample Characteristics

We utilize two primary sources of data: a baseline survey and a follow-up survey. The baseline survey was implemented when the individuals visited EFE for the first time to apply for the program. We built it into EFE’s normal intake form and collected additional data about the trainees’ education, family background, and employment history. The follow-up survey was implemented about 18 months after the randomization of each cohort. In the follow-up survey, we collected information about their current job as well as the first job they took after the training. For each job, we asked about the dates they started (and if they left the job, the date in which they did so), as well as information about the nature of the job.

We use questions about the jobs they have now (18 months after randomization) and the jobs they got right after they were randomized into the program. We consider the former “long-term outcomes” and the latter “short-term outcomes”. The short-term outcomes will be subject to recall bias, but previous research has shown that this potential bias is small on outcomes as salient as these on this time-frame (De Nicola and Gin\’e (2014)). Since we are comparing responses between treatment and control, the recall bias can also “cancel-out” as long as recall is not worse differentially by treatment, which we think is a reasonable assumption, but something we cannot rule out.

Table 1 reports baseline summary statistics about those who entered the randomized sample. The total sample is 985 individuals. On average, trainees in the control group are 24 years old, 73% are women, they graduated from college about 2 years prior to the training, and 96% were single. Two-thirds had a father with post-secondary education, and about half had a mother with post-secondary education. Only 16% were working at the time of the baseline, and about a third of the sample has taken job training before. There are no significant differences between the randomized groups, with a joint test of significance reported in the penultimate row of Panel A.^[7]

We were able to locate and interview 82% of the sample for the follow-up survey. 9 shows that there is no differential attrition by treatment status. We also regress an indicator for attrition on our baseline characteristics and an interaction of those characteristics with treatment and find that a joint test of significance shows no evidence of differential attrition by baseline characteristics.

Analysis & Results

Thanks to the randomization, our econometric analysis is relatively straightforward. To assess the average impact of each training arm on outcomes relative to control, we use the following equation:

which compares all three treatment arms separately to the control group. $Y_{i}$ is the outcome of interest (e.g., income), ${T}$ is an indicator that equals 1 if the individual was in the soft, technical or mixed skills training respectively, and $Y_{0_{DPL}}$ represents the set of baseline controls chosen using the double post-lasso procedure outlined in Belloni et al. (2014).^[8] The $\delta_{C}$ are randomization cohort fixed effects. In all of our analyses, we will be estimating and interpreting “intent to treat” effects, where we compare everyone in treatment to everyone in control whether or not they have received the training.

Our primary outcomes are employment and income. We also consider how details of the job differ in an effort to better understand the mechanisms through which any differences in the primary outcomes may occur. This includes indices for job conditions and job satisfaction as well as an indicator for the occupation being “high-skill”.^[9] We also consider impacts on the beliefs and future plans of the participants.

Testing for Complementarity

To test for complementarity, we follow the method outlined in Mbiti et al. (2019). In their analysis, they have a treatment group ”A” that provides inputs to schools, and a treatment group “B" that provides incentives to teachers, and a group “C" that receives both of these things. They are testing if “C>A+B”, and if so they present this as evidence of the complementary nature of interventions, i.e., A & B being worth more than the sum of their parts. To do so they simply run a test of whether the coefficient on group C minus the coefficients on group A & B is greater than 0. We will implement this test on our data (i.e., $\beta_3>\beta_1+\beta_2$).

Our setting is different from that in Mbiti et al. (2019) in a critical way - our combined group does not receive both A & B, they only receive half of A and half of B. This is because we do not extend the length of the training to provide the combined group with the additional time needed to learn all of the things that are included in those modules. Hence, the more appropriate test for complementarity in our case is if “C>0.5*A+0.5*B". We check this by running a simple test to check if the coefficient on C minus half of the coefficients on A & B is equal to 0 (i.e., $\beta_3>0.5*\beta_1+0.5*\beta_2$).

Take-up of Training

Panel B of Table 1 reports on training attendance, the experiment’s take-up or “first stage”. Column 1 reports the control group mean, and Columns 2-4 show the average difference between the control group and each of the treatment groups. On average, 90% of the treatment groups report attending the EFE training, compared with 14% of the control group, giving us a very strong first stage.^[10] Looking across the different treatment arms, we find that 91% of those randomized into the soft skills group attended the training, compared with 89% of the technical group and 84% of the mixed group.

It is not unusual for some of the control group members to receive some support in randomized experiments (Duflo et al. (2007)). What is key is that the difference between treatment and control is large and exogenous, which our randomization ensures. Furthermore, we take the conservative approach of considering only the intention-to-treat estimates, and we do not scale up our estimates by the difference in training rates as done in standard instrumental variable approaches (i.e., local average treatment effects, treatment on the treated estimates).

Furthermore, 45% of the control group report attending another training program during the time between randomization and the follow up survey. Those in the treatment groups also report attending other training courses. In the bottom row, we see that 51% of the control group participated in some training since the randomization, including the training from EFE, compared with 90-97% of the treatment groups. It is clear that both the control and treatment groups were actively working to improve their skills beyond the EFE training. This fact does not threaten the identification of our estimates, however, because our intervention is still leading to a large exogenous increase in the likelihood that individuals in treatment get training relative to the control group. This means that we have a strong first stage both for getting EFE’s training and for getting any training.

However, while identification of the estimates is not threatened, their interpretation is affected. The treatment effect is diluted due to the control group receiving some training as well. Our intention-to-treat estimates may therefore be an underestimate of the true value of the EFE training. But we do not have data on the content and duration of the outside training received by any group, and this could bias our results in ways we cannot account for, which is a limitation of our analysis.

Ideally, we would be able to directly assess the quality of the EFE training by collecting data on the actual skills of the participants, both before and after the program. However, measuring soft skills is difficult (Devedzic et al. (2018)). In their study, Barrera-Osorio et al. (2023) include self-reported measures of things like work ethic and interpersonal skills, but there is nothing to validate these measures. Instead, during the follow-up survey, we asked the participants if they thought the training was worthwhile and if in retrospect they would have taken it and paid 250EGP for it. About 90% of those in all three groups report that they thought the training was worthwhile, a high level of satisfaction.

Short-Term Outcomes

Both our short-term and long-term outcomes are measured in the follow-up survey conducted about 18 months after the randomization. The short-term outcomes refer to the participants’ answers about their recalled experiences in the few months following the randomization.^[11] Panel A in Table 2 reports results on our primary outcomes in the short term. Column 1 reports the control group mean and columns 2-4 show the average estimated impact of each treatment arm relative to control. While 39% of the control group had gotten a job within 3 months of the randomization, this increases by 10 percentage points for the soft-skills treatment and 14 percentage points for the technical and mixed skills groups. One year after the randomization, 67% of the control group have worked in a job, while there is a 10, 12, and 12 percentage point increase respectively for the three treatment groups.

Columns 5-7 show $p$-values for the differences across the three treatment arms. While we are interested in the general effectiveness of training, we are primarily interested in whether and how the content of the training matters. This tests the recently popular notion that soft skills training may be the answer to increasing the returns to vocational training (Adhvaryu et al. (2021); Groh et al. (2016); Barrera-Osorio et al. (2023); Acevedo et al. (2020)). For the employment outcomes, there is no significant difference between the three treatment arms; all increased employment relative to control by a similar amount.

Looking at our second primary outcome, monthly income in the first job, none of the three treatment arms significantly increased income relative to control, although the point estimates for the technical and mixed groups are quite large (an 11-12% increase relative to control). Looking at columns 5-7, there is strong evidence that the technical and mixed skills groups are economically and statistically different from the soft-skills arm. So while the soft skills group does benefit (along with the other treatments) by finding employment more quickly than the control group, the other two treatments lead to higher incomes.

In Panel B, we consider the details of the jobs to try to better understand the differences in income.^[12] On working hours, the soft skills group did significantly worse than the control and other treatment groups, suggesting that the better performance on income of the technical and mixed treatments is coming through increased working hours. We see no differences across treatments in whether the occupation was high-skill. However, there are significant differences in the proportion of jobs that require English. In the soft skills group, 43% of participants had a first job that required English, compared with 52% for the technical skills group. This is in line with the fact that the technical skills group received the most English-language training of the three groups. We also see some evidence that the technical skills group is most satisfied of the three treatment arms with their jobs.

We thus have three main results on short-term employment. First, the training program seems to have helped people find jobs more quickly than the control group, although there is no evidence that the jobs they found were better or higher-paying. Second, there are large and important differences between treatments, meaning that the content of the curriculum matters. The soft skills training seems to have been the least effective of the three treatments, leading to jobs that look similar (or possibly even worse) than those the control group found. Third, the technical and mixed groups saw higher first-job incomes than the soft skills group, and at least some of this seems related to accessing jobs that require English, which may be at higher-quality establishments (though there is no difference in occupation type). Since these groups received English-language training, and the technical group received the most, it seems the training programs successfully taught the technical skills they were designed to teach. Requiring knowledge of English may be an important and understudied dimension of high-quality jobs in developing countries (Azam et al. (2013)).

Longer-Term Outcomes

Next we consider the jobs held by the sample at the time of our follow-up survey (about 18 months after the randomization). We ask about all the jobs they had since randomization as well as more detailed information about their current job including income, hours, and job conditions. The results are presented in Table 3.

In Panel A, we again consider our two primary outcomes, employment and income. We find that on average the EFE training program seems to have had little long-term benefit relative to the control group. For example, the short-term employment benefits of training fade in the longer term, with 90% of both the control and treatment groups having worked since randomization, and the probability of working in the past month shows a similar pattern. We also see no average benefit of training on monthly income. This lack of overall effect is in line with much of the previous literature (Card et al. (2018)).

Our primary interest in this paper, however, is on whether and how the content of the training matters. Thus, our main results in columns 5-7 compare the three treatment arms to each other. Panel A of Table 3 shows that employment rates are insignificantly higher for the technical and mixed groups (relative to the soft skills group), although the mixed-skill group has had 0.19 more jobs since randomization.

When we look at monthly income, we again see an important difference between treatments. Those in the mixed treatment group increase their income by 240 EGP, a statistically significant 27% increase relative to the soft skills group. The dynamics for the soft and technical groups -- technical does better than soft in the short term, but they are similar in the longer term -- are similar to the findings in Barrera-Osorio et al. (2023).

This difference in income between the soft skills group and the mixed training group is striking, and we explore the potential mechanisms behind this difference in Panel B. The mixed group works 3.5 (12%) more hours per week than the soft skills group, and their wages are 4% higher. The mixed group is both working more and getting paid a higher wage than the soft skills group. While these individual differences are not statistically significant, they add up to the significant difference in income.

Looking at the occupations that people hold, we find that about half of the jobs are high-skilled jobs, and there are no significant differences across treatment groups. Different training programs are not getting people into occupations that are considered higher-skilled. On the other hand, those in the mixed-skill group are 10 percentage points (almost 20%) more likely than the soft skills group to be in jobs that require English skills, and they also report better job conditions. Putting these results together, the mixed group is obtaining better jobs within occupation -- perhaps at higher-quality employers. They are doing jobs that require more language skills. This is important since the returns to English in this context are seen as high, as in many developing country contexts (Azam et al. (2013)).

Interestingly, those in the technical group end up being the most likely to own their own businesses. One interpretation of this is that they learned the skills they needed to succeed in the market, but did not learn the skills needed to succeed in settings where they need to interact with co-workers and managers, many of the skills included in the soft-skills modules.

Taking together the results from Panels A and B, it appears that the mixed-skill group has obtained higher-quality jobs than the other treatment groups by climbing the job ladder more since randomization (i.e., they have moved up to better jobs over time, as evidenced by having more distinct jobs since randomization relative to the other groups). While the technical skills group was most likely to be in jobs that require English in the short term -- consistent with receiving the most English-language training -- the mixed group has overtaken them in the longer term. This suggests that soft and technical skills may be complementary, which we explore in more detail in the next section.

Complementarity

We now explicitly test for skill complementarity using the strategy outlined in Mbiti et al. (2019). We are primarily comparing if the coefficient on the combined group is larger than half of the coefficients on the other treatments (i.e., $\beta_3>0.5*\beta_1+0.5*\beta_2$). This is because those in our combined group spend only half as much time on each of the types of training as in the more specialized treatment groups. We will also check a more stringent test for complementarity that assumes that individuals received the full specialized treatment (i.e., $\beta_3>\beta_1+\beta_2$).^[13]

The results are found in Table 4. We find evidence of skill complementarity for the number of distinct jobs since randomization and for current income. In these cases, the mixed group outcome is more than the “sum of its parts”, since it is a half part of soft skills and a half part of technical skills.

In the bottom row, we consider whether the mixed group did better than the combined impact of the soft and technical groups. This is a stronger test of complementarity and assumes that an individual can get the full effect of each treatment in about half the time. In this case, we find strong evidence of complementarity in income. That is, there is no income benefit to the content of each individual training, but when they are mixed, the training can yield positive results. This is a remarkable and surprising finding. Without strong skill complementarity, one would expect that combining half of two ineffective trainings would only produce another ineffective training. Instead, we find that the mixed training is much better than the other two treatments.

Our results provide an interesting and nuanced contribution to the literature on the effectiveness of training programs. These programs have widely varying returns (Card et al. (2010); Card et al. (2018)), and our findings suggest that some of this could be due to differences in the types of skills taught in the program. However, it is not as simple as teaching soft skills instead of technical skills. Instead, finding the right combination of soft and technical skills seems to be key to maximizing the effectiveness of training for improving longer-term outcomes. Further research is needed on the optimal combinations and exactly how soft and technical skills complement each other.

As noted earlier, we cannot test directly whether the training programs affected the participants’ actual skills, since these (especially soft skills) are difficult to measure well. But we can say that our results are fully consistent with the technical and soft skills trainings being complementary.

Impacts on Additional Outcomes

In Table 5, we consider data we collected on health, mental health, behaviors, attitudes, and expectations for the future. Most of these effects are insignificant, but we do find some evidence that the training changed people’s short-term plans and expectations. Those in the training groups are 7-13 percentage points (about 27%) less likely to report that they plan to get married in the next year, relative to control. However, they are not less likely to plan to get married within three years, and their target number of children is also similar to control. This suggests that the program has delayed, but not fundamentally changed, the trainees’ expectations about family and marriage.

Given the paucity of significant effects on these additional outcomes, we do not want to make too much of the marriage result. But it is possible that a delay in seeking marriage, combined with the increased probability of searching for work (as seen in Table 3), could have longer-term benefits than we measure here. We see no overall effects on reported physical or mental health, or attitudes towards marriage and gender roles in society. The components of these indices are found in 12.

Discussion and Conclusion

Using an RCT in Egypt to provide training to unemployed and underemployed college graduates, we show compelling evidence that soft and technical skills training are complementary, particularly in the longer term. The mixed-skills group, which received half of the soft and half of the technical training, achieved higher incomes and more working hours than the other groups. These results are consistent with a story where the training leads to more climbing of the job ladder by participants - not to different occupations, but to jobs that require English and have better overall income and conditions.^[14] All three treatment groups are more likely to be actively looking for work relative to the control group, but it is those in the mixed group who are successful in turning that effort into better outcomes.

The complementarities between soft and technical skills training have important policy implications. While vocational training often has only modest returns (Card et al. (2010)), some have suggested that teaching soft skills is the key to improving these programs (Adhvaryu et al. (2021); Groh et al. (2016); Barrera-Osorio et al. (2023); Acevedo et al. (2020)). Our results tell a more nuanced story. In our setting, soft skills training on its own seems to have no lasting benefit. However, when combined with technical skills, it can lead to long-term success. Though our results come only from the Egyptian context, they suggest that vocational training programs should teach a mix of soft skills and technical skills and that English language training should be included in cases where there is a potential market premium for it.

Our results also show the need for more research on this topic. We sometimes find slightly negative effects of the soft-only and technical-only treatments, and we lack the data to understand these estimates. It seems that these more focused trainings spent too much time teaching certain skills and that a more diversified training program is better. Knowing specifically which skills should be taught and in what proportions would be a valuable direction for future research.

The soft vs. technical skills debate is incomplete without a full accounting of the ways in which they are complementary. Our study, with college graduates as our subjects, is in a context where soft skills are likely to be the binding constraint. A similar study with a less-educated sample would be insightful. Finding ways to accurately measure soft skills in the field can also help our understanding of the impacts of these programs. We find some evidence that the complementarities allow people to climb to better jobs more quickly. Tracing the careers of trainees in more detail is also necessary to understand exactly how and why soft and technical skills work together. Testing these training strategies in different contexts would help assess if our results generalize to other settings and samples.

Tables

Online Appendix

Notes

1. Here we use vocational training to refer to short-term training focused on giving people skills for employment, as the term is used in papers such as Barrera-Osorio et al. (2023) and Acevedo et al. (2020). There is also a large literature on vocational secondary education, such as vocational/career tracks in high school (Silliman and Virtanen (2022); Brunner et al. (2021)). ↩

2. The technical skills treatment in Barrera-Osorio et al. (2023) is 100 hours of technical skills and 60 hours of soft skills (63% technical), and the soft skills treatment is the reverse (63% soft). Our technical and soft treatments are more dissimilar, allowing us to better test their impacts against each other and against the mixed treatment. ↩

3. We initially disagreed with EFE about whether “labor law” should be called soft or technical. Given that it is only 3 of the 120 hours in soft skills, the classification did not ultimately matter. ↩

4. While some use the terms “soft” and “social” skills interchangeably, Deming (2017) notes that social skills are really a subset of soft skills. There are soft skills that do not involve how to work with others, such as punctuality and grooming. Economists have also used the term “noncognitive” to refer to soft skills (Cunha et al. (2010)). ↩

5. EFE’s mission is to help train university graduates who cannot afford to complement their education with additional training that EFE sees as important to success in the labor market. This is why their inclusion criteria are based on education and perceived financial need. This decision means that our estimates are only valid for individuals of this type, and how much our results would generalize to a different sample is an open question. ↩

6. Osman and Speer (2023) report on an experiment that focused on how to most effectively recruit individuals to job training programs. None of the participants in the program we study in this paper were part of the sample from the other experiment. ↩

7. Some individuals were included in the randomization but did not fill out a baseline survey. For those individuals, we recover time-invariant variables (e.g., age) from our follow-up surveys, and for other variables, we drop the observations from the regression where that variable is an outcome (e.g., balance on education). For regressions where they are used as an independent variable (e.g., for the test of joint significance), we include an indicator for the missing observations. ↩

8. We use the double post-lasso procedure because it provides us with a disciplined way to choose control variables and improves power relative to a simple ANCOVA specification. We include all of our baseline data as options for the lasso. It does not often “choose” any controls, but when it does, the two main controls are a binary for female and baseline English speaking ability. ↩

9. Breakdowns of the components of the conditions and satisfaction indices can be found in 10. To create the index, we standardize each component, add them together, and then standardize the sum (Kling et al. (2007)). We define “high-skill” occupations by categorizing the list of jobs. A list of all job categories and their designation can be found in 11. ↩

10. As part of the experiment, we told the control group at the time of randomization that we could not provide them training now, but if they wanted to get trained in a year they would have an opportunity to do so. This was because the training organization did not want to deny service indefinitely to those who wanted it. Hence, the 14% of control who took the training could be from individuals who took it one year after randomization. Data on training attendance are self-reports from the follow-up survey we conducted. We do not have administrative data on training attendance due to turnover at the training organization. ↩

11. Since we collected data at only one point in time, our short-term outcomes will be subject to potential recall bias. We think there are three reasons why these data are nonetheless informative. First, thanks to the randomization and our cohort fixed-effects, we are comparing answers from people with the exact same amount of recall, and so a portion of the bias will “cancel out” in our analysis. Second, the literature has shown that recall errors over the time frame we are working on are not severe (De Nicola and Gin\’e (2014)). Third, much literature has shown that retrospective data of this kind provides great value, e.g., in migration (Hamory et al. (2021); Beegle et al. (2011)). ↩

12. We estimate impacts on the conditional wage, i.e., the wage paid for those who are working. For this to be a valid comparison, we must assume that the selection of “who” is working is similar. Since we find similar employment rates across the treatment groups we think this is a reasonable assumption. ↩

13. It is not necessary to find positive impacts of the main treatments relative to control to provide evidence of complementarity. It is entirely possible that an intervention could have a negative effect on outcomes relative to no intervention (for example, by not being worth the opportunity cost of time, which has been shown to be true in other training interventions), but a combination of the two interventions works well and is worth more than each of the components individually. Imagine a stylized example where there were two different kinds of training for a research assistant, one that focused on the importance of prioritizing amongst different tasks, and a second that focused on generating well-documented code. The outcome of interest is their ability to complete a coding task in one hour. Imagine that the RA that received no training scored a 75/100 on the task, the one that got the “prioritizing” training scored a 60/100 (because they spent too much time planning), the one that got the “documenting” training scored a 60/100 (because they spent too much time documenting), and the one that got both scored an 80/100. We think that this is an example where there is evidence of training complementarity. Since the combined treatment did better than the combined impacts of the two separate treatments, it is clear that together the combined training is more than the sum of its parts. And indeed, the results from our experiment are similar to this stylized example. ↩

14. By job ladder, we mean we mean the phenomenon where workers move to higher quality firms through job-to-job quits as in (Moscarini and Postel-Vinay (2018)). ↩

References

• Adhvaryu, Achyuta and Kala, Namrata and Nyshadham, Anant (2021). Returns to on-the-job soft skills training.
• Assaad, Ragui and Krafft, Caroline (2013). The Egypt labor market panel survey: introducing the 2012 round. IZA Journal of Labor \& Development.
• Azam, Mehtabul and Chin, Aimee and Prakash, Nishith (2013). The returns to English-language skills in India. Economic Development and Cultural Change.
• Barrera-Osorio, Felipe and Kugler, Adriana D and Silliman, Mikko I (2023). Hard and soft skills in vocational training: Experimental evidence from Colombia. World Bank Economic Review.
• Beegle, Kathleen and De Weerdt, Joachim and Dercon, Stefan (2011). Migration and economic mobility in Tanzania: Evidence from a tracking survey. Review of Economics and Statistics.
• Belloni, Alexandre and Chernozhukov, Victor and Hansen, Christian (2014). Inference on treatment effects after selection among high-dimensional controls. The Review of Economic Studies.
• Brunner, Eric J and Dougherty, Shaun M and Ross, Stephen L (2021). The effects of career and technical education: Evidence from the Connecticut Technical High School System. Review of Economics and Statistics.
• Card, David and Kluve, Jochen and Weber, Andrea (2010). Active labour market policy evaluations: A meta-analysis. The Economic Journal.
• Card, David and Kluve, Jochen and Weber, Andrea (2018). What works? A meta analysis of recent active labor market program evaluations. Journal of the European Economic Association.
• Cunha, Flavio and Heckman, James J and Lochner, Lance and Masterov, Dimitriy V (2006). Interpreting the evidence on life cycle skill formation. Handbook of the Economics of Education.
• Cunha, Flavio and Heckman, James J and Schennach, Susanne M (2010). Estimating the technology of cognitive and noncognitive skill formation. Econometrica.
• De Nicola, Francesca and Gin\'e, Xavier (2014). How accurate are recall data? Evidence from coastal India. Journal of Development Economics.
• Deming, David J (2017). The growing importance of social skills in the labor market. The Quarterly Journal of Economics.
• Devedzic, Vladan and Tomic, Bojan and Jovanovic, Jelena and Kelly, Matthew and Milikic, Nikola and Dimitrijevic, Sonja and Djuric, Dragan and Sevarac, Zoran (2018). Metrics for students’ soft skills. Applied Measurement in Education.
• Dixon, Jami and Belnap, Cody and Albrecht, Chad and Lee, Konrad (2010). The importance of soft skills. Corporate finance review.
• Duflo, Esther and Glennerster, Rachel and Kremer, Michael (2007). Using randomization in development economics research: A toolkit. Handbook of development economics.
• Elbadawy, Asmaa (2015). Education in Egypt: Improvements in attainment, problems with quality and inequality. The Egyptian labor market in an era of revolution.
• Groh, Matthew and Krishnan, Nandini and McKenzie, David and Vishwanath, Tara (2016). Do wage subsidies provide a stepping-stone to employment for recent college graduates? Evidence from a randomized experiment in Jordan. Review of Economics and Statistics.
• Hamory, Joan and Kleemans, Marieke and Li, Nicholas Y and Miguel, Edward (2021). Reevaluating agricultural productivity gaps with longitudinal microdata. Journal of the European Economic Association.
• Hanushek, Eric A and Schwerdt, Guido and Woessmann, Ludger and Zhang, Lei (2017). General education, vocational education, and labor-market outcomes over the lifecycle. Journal of human resources.
• Heckman, James J and Kautz, Tim (2012). Hard evidence on soft skills. Labour economics.
• Kling, Jeffrey R and Liebman, Jeffrey B and Katz, Lawrence F (2007). Experimental analysis of neighborhood effects. Econometrica.
• Krueger, Dirk and Kumar, Krishna B (2004). Skill-specific rather than general education: a reason for US--Europe growth differences?. Journal of economic growth.
• Mbiti, Isaac and Muralidharan, Karthik and Romero, Mauricio and Schipper, Youdi and Manda, Constantine and Rajani, Rakesh (2019). Inputs, incentives, and complementarities in education: Experimental evidence from Tanzania. The Quarterly Journal of Economics.
• Moscarini, Giuseppe and Postel-Vinay, Fabien (2018). The cyclical job ladder. Annual Review of Economics.
• Osman, Adam and Speer, Jamin D and Weaver, Andrew (2022). Connections, Referrals, and Hiring Outcomes: Evidence from an Egyptian Establishment Survey. Journal of Economic Behavior and Organization.
• Osman, Adam and Speer, Jamin D. (2023). Stigma and Take-Up of Labor Market Assistance: Evidence from Three Experiments.
• Paloma Acevedo and Guillermo Cruces and Paul Gertler and Sebastian Martinez (2020). How Vocational Education Made Females Better Off but Left Males Behind. Labour Economics.
• Schulz, Bernd (2008). The Importance of Soft Skills: Education beyond Academic Knowledge.. NAWA Journal of Language \& Communication.
• Silliman, Mikko and Virtanen, Hanna (2022). Labor market returns to vocational secondary education. American Economic Journal: Applied Economics.
• Weinberger, Catherine J (2014). The increasing complementarity between cognitive and social skills. Review of Economics and Statistics.
• \firstsecondILO (2016). Employent by Sex and Age -- ILO Modelled Estimates.