Policy Research Working Paper 8141 The Role of Skills and Gender Norms in Sector Switches Experimental Evidence from a Job Training Program in Nigeria Kevin Croke Markus Goldstein Alaka Holla Gender Cross Cutting Solution Area July 2017 Policy Research Working Paper 8141 Abstract Industrialization and structural change entails shifting those with sector-relevant skills, and training magnified workers from low skill to high skill occupations, and in the skills premium in switching. Switches were also higher emerging economies a number of constraints may impede in some cities, despite large improvements in skills in all sectoral switches among workers, including skill and spa- cities. Women who were implicitly biased against associ- tial mismatches, and social norms related to gender in the ating women with professional attributes were two times workplace. This study uses a job training experiment across more likely to switch into ICT than unbiased women, five cities in Nigeria to estimate the overall effect of training suggesting that training helped overcome self-defeating on sectoral switches into the information and communica- social norms among female applicants of the program. tions technology and business process outsourcing sector, These results suggest that training can be an effective strat- and to examine the role of various factors that might con- egy for inducing sector switches and overcoming social strain switching. After two years the treatment group was norms that hamper female mobility in the labor market, 26 percent more likely to work in the ICT-enabled service but that addressing spatial mismatches in labor demand sector, although they were no more likely to be employed and labor supply may require additional interventions. than the control group. Sector switches were higher among This paper is a product of the Gender Cross Cutting Solution Area. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at ckroke@worldbank.org, mgoldstein@worldbank.org, and aholla@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team The Role of Skills and Gender Norms in Sector Switches: Experimental Evidence from a Job Training Program in Nigeria Kevin Croke, Markus Goldstein, and Alaka Holla* Updated January 2021 JEL classification: J24, D91 Keywords: Formal training programs, occupational mobility, active labor market programs, randomized experiment * Croke: Harvard T. H. Chan School of Public Health,1818 H St. NW, Washington DC 20433 (e-mail: kcroke@hsph.harvard.edu). Goldstein: World Bank and BREAD, 1818 H St. NW, Washington DC 20433 (e-mail: mgoldstein@worldbank.org). Holla: World Bank, 1818 H St. NW, Washington DC 20433 (e-mail: aholla@worldbank.org). We thank Maria Elena Garcia Mora and Harsha Thirumurthy for work on the initial stages of the evaluation. We also thank our operational collaborators, in particular Toks Fayomi, Marito Garcia, Peter Materu and Anubha Verma. Tigist Ketema and Daniel Kirkwood provided valuable research assistance. Project Implicit provided critical support in implementing the implicit association tests. This product is an output of the Africa Gender Innovation Lab of the World Bank. Support from the Umbrella Facility for Gender Equality is gratefully acknowledged. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Introduction Industrialization entails shifting workers from low skill to high skill occupations. While economic development in many currently high- and middle- income countries followed a path away from agriculture towards manufacturing, many emerging economies today aim for service-led growth. These economic transformations are happening in contexts where a number of constraints may impede efficient allocation of human capital across sectors, including skill and spatial mismatches and social norms related to gender in the workplace (Campos et al., 2015; Jensen, 2012; and Bryan et al., 2014). In this paper, we study the extent to which these constraints affect sector switching into the information and communication technology (ICT) industry through a randomized control trial that offered training for certification in sector-relevant skills in urban Nigeria. The training program targeted recent university graduates, seeking to improve their proficiency in basic software packages used in the business process outsourcing (BPO) sector as well as their oral and written communication skills, in order to prepare them for work in the emerging ICT and BPO sectors. Before the program, a general deficit of skills demanded by firms was considered a large handicap for economic growth in Nigeria (Treichel, 2010). Indeed, a comparison between India and Nigeria focused on the ICT industry found a sizable skills disadvantage among Nigerian youth (World Bank, 2012). At the same time, individuals with sector-relevant skills could not signal their abilities, as there was no certification exam recognized by firms in the emerging sector. More generally, youth unemployment was and still is of increasing concern to policy makers, not just in Nigeria but across the world, especially in countries in the Middle East and Sub-Saharan Africa currently experiencing a significant youth bulge (Filmer and Fox 2014). Bleak job prospects are not limited to populations with low levels of formal education. Indeed, for low- income countries in Sub-Saharan Africa, the unemployment rate is higher among individuals with tertiary, vocational, or university education (at 18.8 percent), than it is for other education groups (African Economic Outlook 2012). Many youth employment interventions globally have focused on improving skills of the unemployed, under the assumption that there may be poor match of skills between young people emerging from the formal education system and the profiles sought after by the private sector. 2 Experimental evidence for the effectiveness of this strategy in low-income countries, however, is limited, as noted by Betcherman et al. (2004). More recently, Card et al. (2015) aggregated over 200 econometric evaluations of active labor market programs, finding no average effects on employment in the short run but positive effects over longer time horizons, while McKenzie (2017) finds modest impacts of vocational job training across low- and middle-income countries. Taking a broader look at labor market and entrepreneurship programs in low-income countries, Blattman and Ralston (2015) argue that “it is hard to find a skills training program that passes a simple cost- benefit test.” In the case of our experiment, we find modest effects of the program on ICT employment (1.7 percentage points from a baseline of 6.4 percent) and no effects on employment or earnings more generally, suggesting that the job training served to shift workers to the sector targeted by the program. Heterogenous treatment impacts suggest that to some extent the program helped overcome skills mismatch present in the nascent industry, as individuals with higher scores on a baseline assessment that overlapped with the certification exam that firms began to use were much more likely to benefit from the program. Impacts also varied across cities, which suggests the presence of spatial mismatches in labor supply and labor demand, consistent with the literature that finds high returns to programs that induce internal migration for work (Jensen, 2012 and Bryan et al., 2014). After the program, the share of the treatment group meeting international benchmark scores for entry-level positions grew substantially in all cities, but the training generated varying amounts of sector switching across cities. At baseline, we also measured applicants’ attitudes and biases related to gender norms. At the time of the experiment, the ICT sector was male dominated, with government figures indicating that 67 percent of workers in the information services sector in 2010 were men (National Bureau of Statistics 2010). 1 Additional heterogenous treatment impacts suggest that training can play a role in changing these norms and aspirations. While structural features of the economy and outright discrimination are likely dominant reasons for women’s lower labor market earnings, additional disadvantages may stem from beliefs that women themselves hold about their own potential. Women who hold gender-based stereotypes with respect to their mathematical abilities, for 1 There is no direct mapping between the goals of the training program and occupational categories in government statistics. All other possible categories (office support, telecommunications) are also male dominated. 3 example, perform worse in math exams and demonstrate lower interest in mathematical careers (Spencer, Steele, and Quinn 1999; Kiefer and Sekaquaptewa 2006). In the Netherlands, even when boys and girls may demonstrate similar levels of academic ability, boys tend to choose more academically prestigious tracks (Buser, Niederle, and Oosterbeek 2014). Internalized beliefs and biases, however, are difficult to measure using standard survey techniques either because individuals are unaware of the biases they harbor, or because self-reported biases often conform to what the respondent thinks the listener would like to hear. In this study, we use a tool developed by psychologists to measure implicit biases, the Implicit Association Test (Greenwald et al, 1998). For women who were implicitly biased against associating women with professional attributes at baseline, the likelihood that the program induced switching into the ICT sector was more than two times as large than for unbiased women. This suggests that training programs have the potential to overcome self-defeating biases that reinforce occupational segregation, even when they do not explicitly set out to do so. In this particular case, we hypothesize that training to work in the ICT sector changed women’s ability to imagine themselves as professionals in ICT. While job training programs typically focus primarily on the transfer of skills, our results are consistent with a small empirical literature which suggests that job training also offers a potential opportunity to shift beliefs and norms about the appropriate sectors for men and women to work in, and thus reduce occupational segregation (Campos et. al. 2015). Nopo, Robles and Saavedra (2008) examine the impact of a vocational training program in Peru which explicitly sought to reduce occupational segregation by encouraging females to enter male-dominated occupations. Using propensity score matching, the authors find that 18 months after the end of training, women were 15 percent more likely to be employed and had a 93 percent increase in earnings, while the program reduced occupational segregation by 30 percent. Globally, occupational segregation is a phenomenon common to both high- and low-income settings. Blau and Kahn (2016), for example, find that observed gender differences in occupations and industries are the most important factors underlying the gender wage gap in the United States. In Sri Lanka, once industry is taken into account, female ownership loses its power to explain gender performance differences among small business owners (de Mel et al., 2009). More generally, the World Development Report, using data from 33 low and middle-income countries 4 and 14 high income countries, shows that this segregation accounts for 10-50 percent of the wage gap (World Development Report 2012, p. 206). The next section provides more detail on the Nigerian context and describes the job training intervention, including the experimental allocation of training slots. Section 3 presents the data sources used for evaluating the program and outlines the empirical strategy for estimating treatment impacts and gauging the importance of various labor market frictions for sectoral switching. Section 4 presents results and robustness checks, and Section 5 concludes with a discussion of their relevance for policy. Country context and program background In 2010, ICT was an emerging industry in Nigeria. Even though agriculture was the largest aggregate contributor to the country’s high economic growth rate in this period, telecommunications was the fastest growing sector (with a growth rate of 30 percent, compared to 7 percent for agriculture). There were approximately 400 small and medium-sized ICT firms in Nigeria at the time; these were mostly call centers catering to the domestic market, primarily large corporations in banking and telecommunications (World Bank 2012). A long-term goal of the government was to serve the large public sector, and to break into the international market for ICT- enabled services (for example, call centers and other forms of business processing outsourcing). They sought to do this by taking advantage of Nigeria’s low labor costs and abundant supply of English speakers, just as emerging economies like India, the Philippines, and the Arab Republic of Egypt had successfully done in the previous decade. Industry consultations suggested that a large skills gap among potential workers prevented the industry from being internationally competitive. In the second half of 2010, the Government of Nigeria implemented the ACCESS (“Assessment of Core Competency for Employability in the Service Sector”) Nigeria program, with the immediate goal of training recent university graduates, equipping them with sufficient skills to work in Nigeria’s ICT sector, and certifying these skills. An initial activity of the program was an assessment of Nigerian university graduates, which found lower competency levels for Nigerians compared to Indians and North Africans in almost all skill categories (for example, voice clarity, ability to use Microsoft Office software, and listening comprehension). Nevertheless, the assessment suggested that the Nigerian university graduate population had strong potential to meet international industry standards. The training was expected 5 to improve skills in three competency areas which are considered “foundational competencies” for employment in the business processing outsourcing (BPO) sector (Eduquity 2012): communication (oral and written), computers, and cognitive skills. The government implemented the ACCESS Nigeria IT job skills training program in five cities in 2012 (Abuja, Enugu, Kaduna, Kano, and Lagos). Outreach efforts through university campuses and via radio advertisements in these cities encouraged applications from university graduates, particularly those finishing their final year of university and those currently doing their year of service in the National Youth Service Corps – a mandatory internship that must be completed after university but prior to formal sector employment. The ACCESS Nigeria program consisted of 85 hours of classroom-based training spread across 10 weeks. At the end of training, program participants could also take an assessment exam which had been recognized by the domestic ICT industry as a form of certification to work in business processing activities. A consulting firm designed a curriculum based on curricula used in other countries (Egypt, India, Jordan, Philippines, Singapore, and Sri Lanka), an initial evaluation of recent university graduates in Lagos, and an assessment examination that had been endorsed by an industry consortium as certification for the sector. The government chose training providers in each of the five cities through a competitive bidding process, and training providers had to transform the curriculum into concrete lesson plans. The chosen training providers were private firms, except in Abuja, where a public sector institution won the bid. In addition to the core competencies, these training providers also were required to cover “soft” skills, such as cultural sensitivity, teamwork, and stress and time management. Figure 1 presents a timeline of all intervention and evaluation activities. The radio advertisements and outreach activities in local universities attracted 3,018 eligible applicants to the program. ACCESS Nigeria funds paid for the training slots, with an average cost of approximately $600 per trainee. From March to April 2011, all applicants were invited to come to training centers to take a computer-based, self-administered assessment exam, the content of which heavily overlapped with the certification exam. Using this data, we generated random numbers and assigned them to applicants within strata defined by scores on the pre-assessment, as well as gender, test center location, and an applicant’s academic status (final year student in university, participant in the National Youth Service Corps, and all others). Within each stratum, the 60 percent of applicants 6 with the highest random numbers were selected for the program. Considerable delays in program implementation led to a lengthy interval between the pre-assessment of applicants (March/April 2011) and the communication of treatment status to applicants (November/December 2011). Training finally commenced in February 2012. We note that this was a politically unstable period in Nigeria, coming shortly after the resolution of a national crisis over the removal of fuel price subsidies and a Boko Haram bombing that killed over 150 people and led to the imposition of a dusk to dawn curfew in the program state of Kano. All training activities were completed by mid- April 2012. Those who completed training could take the assessment again and receive a certificate. All applicants, regardless of their treatment status, could post their resumes on a web- based employment network and attend a job-fair in Lagos in late 2012, where prospective employers could meet with job candidates interested in working in the sector. Figure 2 presents the full experimental design, including compliance rates and sample attrition across rounds using a CONSORT flow diagram. Given the lag between the initial assessment and the final allocation of program slots and the political instability immediately before the start of the program, compliance with treatment assignment was relatively high. According to training records, approximately 54 percent of the treatment group took the offer of training and attended at least one training session. There was no “contamination” or crossover attendance in training by the control group. However, administrative attendance data from training providers shows that average attendance at individual training sessions was far lower, and varied by region. In Lagos, Kano, and Kaduna, for example, training sessions were on average attended by only one-third of those who had been offered treatment, while in Abuja is was over 80 percent, according to training provider records. 2 While gender was not an explicit focus of the program, the study did include consultations with training providers prior to program implementation to determine whether female trainees and job- seekers in the ICT-BPO sectors faced any obstacles in seeking and maintaining employment. As in many countries, gender inequality is a feature of the Nigerian labor market and Nigerian society more generally, even among women with tertiary education. In 2013, for example, one third of women who had obtained a level of education higher than secondary school had not worked at all 2 We do not have aggregate attendance figures for training in Enugu. 7 in the previous 12 months, while this figure was less than one fifth for similarly educated men (Demographic and Health Survey, 2013). Despite their relatively high education, only 53 percent of these women reported that they could decide for themselves how their earnings were used. On the other hand, among these highly educated working women, at least half of them worked in an office setting, with 52.2 percent reporting occupations that could be classified as professional, technical, or managerial. 3 Slightly less than half of men in this category worked in an office job 4. Therefore, this data on representation across occupations suggests that these highly educated female applicants should have had reasonably equitable chances of securing an office job using the program-provided training. Nevertheless, interviews with the target population and training center directors prior to the experiment suggested that women in Nigeria may face higher obstacles in obtaining employment in an office setting, not only because of labor market discrimination, but also because of women’s own confidence in seeking work in the formal sector. To test this hypothesis, the initial assessment taken by all applicants also included questions and tasks designed to measure biases related to female labor force participation, female participation in an office job, and female professionalism. While the government was interested in building the ICT-enabled service sector across all five cities that participated in the experiment, these cities demonstrated very different levels of sector maturity at baseline. Of the approximately 470,000 people across the country employed in the information and communication sector in 2010, 18 percent were in Lagos (Nigeria’s largest city), 3 percent in Kano, 1 percent in Kaduna, 0.8 percent in Enugu, and 0.7 percent in Abuja (National Bureau of Statistics, 2010). Empirical strategy Data The computer-based pre-assessment provided a platform to collect baseline data on the 3,018 applicants of the program. A self-administered questionnaire followed the assessment, with questions on applicants’ socio-economic and demographic backgrounds, education history, and 3 While not all occupations classified as sales or services can be considered office jobs, including these as office jobs would bring the percentage to 87.2 percent. 4 This figure increases to 73 percent when sales or services responses are included as professional, technical, or managerial. 8 labor market experiences and expectations. While not ideal because of potential fatigue and possible inconsistency in how respondents interpreted questions without the aid of survey enumerators, obtaining baseline data through this method solved the logistical and financial obstacles of physically tracking down a potentially scattered and extremely mobile population for face-to-face interviews, as they had been targeted through universities and radio ads and therefore did not necessarily reside in the cities where pre-assessment testing took place. We worked with Project Implicit to include Implicit Association Tests (IATs) in the baseline survey. Social psychologists use IATs to measure an individual’s automatic associations between a social group and a stereotypic attribute (Greenwald, McGhee, and Schwartz 1998; Nosek, Banaji, and Greenwald 2002). Economists have used IATs to measure stereotypes related to gender and occupation (Beaman et al. 2009), race and intelligence (Bertrand et al. 2005), and attitudes towards other ethnic groups (Lowes et al. 2015). The IAT requires sorting of exemplars from four concepts – for example male, female, office, home — using just two response options —for example, left for either male or office and right for either female or home. If it is easier for participants to mentally pair pictures of men with words associated office and women with words associated with the home, then subjects should be able to make these pairings faster than the opposite pairings (women and the office and men and the home). That is, an individual biased by traditional gender roles should be faster in sorting words when the options are {Left: male or office; Right: female or home} than when they are {Left: male or home; Right: female or office}. Because the differences in sorting times are often less than a second, these associations are considered implicit and automatic, or beyond conscious control. Indeed, they often differ from explicitly expressed attitudes. Females, for example, often exhibit stronger implicit attitudes linking males with career and females with family than males, despite reporting weaker explicit attitudes (Nosek, Banaji, and Greenwald 2002). The IATs attached to the assessment and baseline survey consisted of tests measuring the ease of associations between gender and a number of attributes relevant for women’s labor market participation in Nigeria. One test measured associations between gender and the concepts of home and office, a test commonly used in the IAT literature to measure gender-related bias (see, for example, Nosek et al. 2002). A second test was designed for the urban Nigerian context and measured associations between gender and the concepts of office and petty trade, as women might 9 be more associated with less remunerative self-employment activities, such as the sale of phone cards or prepared food, although evidence suggests only a slight increase in the likelihood of being self-employed for women among the urban population with some tertiary education. 5 A final test measured a more subtle distinction but one that is possibly important for applicants already interested in training for the ICT sector: associations between gender and the concepts of professionalism and unprofessionalism. 6 Prior to the sorting tasks, respondents were asked explicit questions about gender and professionalism. Appendix Table 1 presents the basic format of the IATs used during the baseline survey, and Appendix Table 2 lists the wording of the explicit question and the words used to represent home, office, petty trade, professionalism, and unprofessionalism. The endline survey took place by phone between February and April 2014. Enumerators in call centers contacted applicants using the contact information provided during the baseline survey and administered a relatively short end line questionnaire over the phone. They reached 2,733 applicants for a response rate of 91 percent. Training providers also collected attendance data for all trainees and post-assessment scores for applicants who accepted treatment (1,007 individuals), administered at the end of training. Outcomes Our main outcome of interest is employment in the ICT sector, the objective of the program, specifically defined as working in the BPO sector or information and communications technology more generally. During the phone survey, we asked the original baseline participants if they were working and which sector they worked in. We counted anyone who chose response options including information and communication technology, software, computers, call centers, or business process outsourcing as their sector as working in ICT. The training provided skills that could easily transfer to other office jobs outside of ICT-enabled services. To determine whether any observed increase in ICT employment results from an increase in labor market participation or sector switching, we also measure impacts of the training program 5 In Nigeria’s 2011 General Household Survey, 28 percent of urban women with some tertiary education worked in non- agricultural self-employment activities, compared with 21 percent of urban men with similar education. 6 Working with Project Implicit (https://implicit.harvard.edu/implicit/) and local focus groups, we collected a set of words that the target group associated with the relevant concepts (home, office, petty trade, professionalism). 10 on overall employment and reported hours worked. If we find impact on ICT employment without any concomitant increase in labor market participation, this would suggest that the training program induced sectoral switching. During the phone interview, we also asked respondents if they were currently searching for a job and which sector they were targeting. If an increase in job search accompanied an increase in employment or sector switching, this would suggest that the training program increased participants’ mobility in the labor market. We also investigate whether the program increased self-reported earnings, although we note that our measure of earnings is very noisy, as would be expected from two questions on earnings 7 administered over the phone. Finally, Project Implicit provided the scores we used for measuring gender related biases. We measured bias both explicitly (for example, the difference between self-reported rankings of women’s and men’s professionalism) and implicitly through the D-score, a within-participant standardized difference between the sorting times for the different pairings of groups (men and women) with concepts (home and office) (Greenwald et al., 2009). For the specific gender-based IATs implemented at baseline, positive D-scores indicate stronger implicit bias against women. For example, a positive D-score in the gender and professionalism IAT corresponds to a relatively longer time required to associate women with professionalism (and men with unprofessionalism) than men with professionalism (and women with unprofessionalism). We also use indicators of bias defined using ranges of D-scores provided by the Project Implicit team to indicate different levels of bias (none, moderate, strong). Scores between 0 and ±0.15 are considered to indicate little to no bias. Positive deviations greater than 0.15 are considered to indicate pro-male bias, while negative deviations less than -0.15 are considered to indicate pro-female bias. While the program paid for a post-assessment for the entire treatment group, funding constraints did not permit a separate assessment for the control group. Thus, we will not be able to identify whether the training increased industry-recognized skills, as any changes observed in the treatment group could reflect both the impact of training and any learning that the sample might have experienced between the baseline assessment and the training, a period of just under one year and likely a time when individuals in the sample were just entering the labor force. 7 We asked separately about wage earnings and self-employment earnings. 11 Estimation To measure the average impact of the training program, we first estimate intention-to-treat (ITT) effects by regressing employment outcomes on an indicator for treatment offer and all stratifying variables, which include gender, whether the respondent was in National Youth Service Corp (NYSC) or final year of education at the time of application, whether the respondent was above or below the median baseline skills assessment score, and indicators for each initial assessment site, δ: �⃗ + = 0 + 1 + 2 + 3 + 4 + 5 + This specification only makes use of the end line data, since most of the sample was in school or in national service at baseline and hence questions such as sector of employment or salary were not applicable. For a sub-set of variables that are measured at baseline, we can also estimate an ANCOVA specification that controls for baseline realizations of the employment outcomes. To explore the role of potential labor market frictions, we also estimate heterogeneous treatment effects for each stratifying variable. Interpreting the main effects together with the interactions of each of these variables with an indicator for treatment offer can tell us the extent to which the program either reduces or exacerbates any labor market premium associated with gender, skills, internship, and location. Given that we have non-compliance (not all individuals in the treatment group took up the training), we also estimate treatment-on-the-treated effects, in which treatment assignment serves as an instrument for program participation: �⃗ + = 0 + 1 + 2 + 3 + 4 + 5 + �⃗ + = 0 + 1 + 2 + 3 + 4 + 5 + We define program participation in two different ways, although they are highly correlated. The first uses respondents’ self-reports of participation, while the second treats a respondent as a participant only if he or she appears in the attendance records of training providers. Finally, we also use the data from the IATs to measure an additional set of heterogeneous impacts related to biases exhibited at baseline, 12 �⃗ + = 0 + 1 + 2 + 3 ∗ + 4 + 5 + 6 + Finding µ3 > 0 would suggest that applicants exhibiting bias against women at baseline benefit from the program more. Results In this section, we describe the baseline characteristics of the sample and present the results of our empirical specifications. The program appears to have on average increased labor market mobility and augmented the returns to skills. Effects were concentrated in one location and were higher among women who harbored biases against female professionalism at baseline. Descriptive statistics The random assignment of the program produced balance across the treatment and control groups, with no significant differences across 26 covariates, as shown in Table 1. As one might expect, given that college graduates were the target population, the sample appears to be relatively well- off in the Nigerian context. For example, 81 percent of respondents’ mothers and 94 percent of fathers are literate, while the literacy rates overall for Nigerians aged 45-49 are 36 percent for women and 65 percent for men (Demographic Health Survey, Macro International 2014). Appendix Table 3 shows that over 90 percent of participants were re-interviewed at end line, a relatively high response rate among other experimental studies in similar settings that sought to measure labor market outcomes more than one year after a job training program (McKenzie, 2017). 8 There was some differential attrition, as we could complete the end line survey with 92 percent of the treatment group, but only 88 percent of the comparison group responded to the survey. We discuss the implications of this for our results in Section 4 and estimate upper and lower bounds for our main treatment effects. While Figure 2 suggests an overall compliance rate of 54 percent, Table 2 shows some heterogeneity across strata when program take-up is regressed on indicators for each stratum. Individuals who applied during their mandatory internship were close to 10 percentage points less likely to take-up the program, while scoring above the baseline assessment score was associated 8 In fact, among the 12 evaluations of job training programs in low and middle income countries reviewed in McKenzie (2017), only one study had an attrition rate less than 18 percent. 13 with higher take-up. Relative to applicants tested in Abuja, the capital city, applicants tested in Lagos were 38 percentage points less likely to accept a training slot when offered through the experiment. Recall that at this time, Lagos accounted for 18 percent of the country’s employment in the information and communications sector; other providers may have already met demand for training. Similarly, at baseline, scores on the assessment differed across strata. Females scored 0.068 points (or about 0.1 standard deviations) lower, while those during their employment internship scored 0.096 points higher. Relative to Abuja, applicants in Kano and Kaduna scored 0.26 (0.37 standard deviations) and 0.43 (0.62 standard deviations) points lower on the baseline assessment, while applicants from Lagos faired significantly better. To put the magnitude of these scores, which run from 1 to 5, in perspective, most international BPO firms require a score of 4 or above for entry- level jobs. At the time of baseline, only the top 15 percent of applicants would have qualified for entry-level work. Figures 3 to 5 present the basic results of the Implicit Association Tests. 9 In general, both female and male applicants on average exhibited little bias against associating women with professionalism (Figure 3). Female applicants exhibited less bias – both explicit and implicit. These biases were stronger when contrasting the concepts of home and office (Figure 4). Consistent with results from large web-based samples (Nosek, Banaji, and Greenwald 2002), women on average showed a larger implicit bias than men in associating women with the home rather than the office. When it comes to associations between women and petty trade (as opposed to the office), male applicants demonstrated stronger implicit biases than women on average (Figure 5). Basic results While we cannot causally attribute any change in the treatment group’s score on the assessment to the training program (due to the lack of end line skills data for the control group), we do see sizeable changes across the board in the fraction of the treatment group that would meet the international benchmark assessment score of 4 for entry level work (Table 3). While only 16 percent met the benchmark at baseline, after the training, nearly half of the group who took up the 9 Appendix Table 4 checks for differences between the full sample and the sample of baseline respondents who took the Implicit Association Tests. Female applicants, final year students, and those taking the assessment in Kano, Kaduna, and Enugu were less likely to take the tests. 14 training met the benchmark. These gains in skills occurred across all experimental sites and for both males and females. To put these changes in perspective, prior to training, applicants’ assessment scores put them below benchmarks used in the South Asian and North African BPO markets in all assessed skill categories (see Table 8 for a list of these categories). After the training, scores among the treatment group suggest an improvement in skills. In five of these skill categories, those who received training scored above the South Asian and North African benchmarks (Eduquity, 2012). Again, without post-assessment scores among the control group, we cannot determine whether the training program was responsible for this improvement in skills, but the following analysis that takes advantage of the experimental design does suggest positive impacts of the program. Intention-to-treat estimates suggest a moderate average impact of the program on ICT employment (Table 4, Column 1). Two years after training, those offered treatment had a 1.7 percentage point higher likelihood of being employed in the ICT sector, from a base of 6.4 percent, representing a 26 percent increase employment in the sector. This effect is marginally significant (p=0.066), most likely due to the heterogeneity of impacts across groups defined by baseline skills and location (See Section 4). Impacts on other employment variables (Columns 2 and 3) suggest this gain in ICT employment represents only a shift in sectors, as neither overall employment nor hours worked significantly increased in response to the program. With currently available data, it is not possible to isolate a specific explanation for this. The program could have provided skills that did not increase the target population’s general employability but rather just their potential in the ICT sector. It is also possible that firms outside the sector did not recognize the value of the certification exam that trainees took at the end of the course. The estimated impacts on general job search (Column 5), however, do suggest that the program moderately increased (5.4 percentage points) the likelihood that an individual reported that they were currently searching for a job, although this search did not appear to target the ICT sector (Column 4). While total earnings did not increase significantly (Column 6), the magnitude of the estimate is quite large, as is the standard error. Moreover, inverse power coefficients (Andrews, 1989) suggest our study had statistical power sufficient to distinguish only effects larger than 42 percent from a 15 zero increase in earnings. Appendix Tables 5 and 6 show that an ANCOVA specification that controls for baseline values of the employment outcomes and a specification that includes a set of demographic and socio-economic variables do not yield qualitatively different results. In addition to the treatment coefficients, there are several other notable patterns in the data. In all regressions, scoring above the median on the pre-program assessment test is strongly associated with better outcomes, suggesting that the certification assessment does capture skills that are highly valued in the labor market. Applicants that had scored above the median are 90 percent more likely to work in the ICT sector. Scoring above the median is also associated with an approximately 77 percent premium in earnings. Women, however, have a significant disadvantage in the labor market, even in this relatively high skill and high socio-economic status sample. They are much less likely to work in the ICT sector, they earn half as much as men, and they work 28 percent fewer hours. Applicants initially assessed in Lagos also demonstrate a significantly higher likelihood of working in the ICT sector. Treatment-on-the treated results suggest an impact that is twice as high, either when program participation is self-reported or when it is measured by attendance records (Table 5). This should not be surprising since there was no spillover from the treatment group into control group, making the TOT coefficients essentially the ITT coefficients divided by the proportion of the sample that took up treatment (0.55). All other employment outcomes show the same sign and significance as in the ITT specification, although the effects are notably larger. Heterogeneous treatment effects These modest average treatment effects are perhaps not surprising as the program - when offered through the experiment studied in this paper - imposed no selection criteria other than recent or imminent graduation from a university. It is possible that the program may have worked better for some groups more than others, particularly since labor demand may have differed across gender, skill-levels, and labor markets. To check for heterogeneous treatment effects and explore the role of potential sources of labor market frictions, we first interact treatment with each of the variables that were used to stratify applicants when assigning treatment: gender, test center site, educational status (participation in the National Youth Service Corps versus those in their final year of university or those who had 16 already graduated), and whether the respondent was above or below the median pre-training assessment score (Table 6). The results in Panel A Column 1 suggest that the average gains in ICT employment following the training program are statistically indistinguishable for males and females, as is the increase in the likelihood of any job search (Panel A Column 4). However, the program appears to have increased ICT-specific search only for women and to have nearly closed the gender gap in ICT-specific search observed in the control group. In Column 1 of Panel B, we see that the average gains in ICT-sector employment (Table 4) are entirely driven by the applicants who had scored above the median score on the assessment administered at baseline. For this group, the training led to a 4.8 percentage point (s.e. = 0.020) additional increase in employment in this sector, whereas those in the treatment group scoring below the median at baseline could not be statistically distinguished from the control group at the time of the endline survey. Those scoring above the median at baseline also had similar (non) reactions to those scoring below the median when we estimate impacts for general employment, general job search, and ICT-specific search. The treatment interactions with the geographical location of the applicants’ initial assessment sites suggest that the gains in ICT employment were strongest in Abuja (the omitted category), where those offered treatment were 7.7 percentage points significantly more likely to find a job in the ICT sector (Table 6 Panel C Column 1). In the other locations, the training was not as effective in increasing employment in this sector; in Lagos and Enugu the point estimates are lower for the treatment-city interactions, but not statistically different. There are many potential explanations for variation in impacts across cities. Part of the difference could be due to political issues in Nigeria that intervened between baseline and end line. As indicated above, terrorism affected the northern areas of the country during this period, and we can see significantly lower effects for cities located in these zones (Kano and Kaduna). The heterogeneity in program impacts across cities could also result from differences in labor demand. As the coefficients on the state indicators in Panel C Column 1 suggest, compared to the control group in Abuja, the control group in the other cities have a much higher likelihood of finding ICT employment. They also engage in more ICT-specific job search (Panel C Column 2). 17 Thus, labor demand for college-educated workers with ICT-relevant skills may have already been satisfied in these cities. Finally, differences in training quality could also explain the heterogeneity in treatment impact across locations, as a public sector organization won the bid to provide training in Abuja while private sector firms won the bids in the other cities, and reported attendance was notably higher in training sessions in Abuja. We note that such an explanation would be inconsistent with empirical evidence from a training program in Turkey, where Hirschleifer et al (2016) find that private firms were more effective than government training institutes in improving labor market outcomes, but consistent with evidence from France, where Behaghel, Crepon and Gurgand (2013) find strong performance by public providers of job training and search assistance. Table 6 Panel D also shows that applicants applying during their NYSC internship did not benefit any more from the program than final year students, individuals not participating in the internship program, or individuals who had graduated from university much earlier (Table 6 Panel D). Table 7 examines the extent to which gender bias mediates program impact, focusing on the contrasts developed for this study related to professionalism, working outside the home, and working in an office job. We estimate impacts for the female and male samples separately, restricting attention to ICT employment as an outcome. We also examine impact using a measure of explicit bias. Given that applicants who took the Implicit Association Tests were significantly different than those who skipped this component along some dimensions (Appendix Table 4), we estimate the basic ITT specification on the sample of applicants who took the IAT (Table 7, Column 1) using relevant D-scores as our measures of implicit bias. The estimated impact of the program for this sample is statistically indistinguishable from the full sample estimate presented in Column 1 of Table 4. We first see that explicit bias related to female professionalism does not alter treatment impact for either females or males (Columns 2 and 6). While difficult to interpret the units of the D-score, the results suggest large differences between females biased against female professionalism in treatment impact, where a unit increase in the D-score increases treatment impact by 7.8 percentage points (Column 3), more than doubling the average treatment effect. 18 In Table 6 (Columns 2 and 4), we saw that the job training program increased any kind of job search for both male and female applicants and increased ICT job search more for female applicants than male applicants. Although not reported here because of space limitations (results available upon request), treatment impacts on general search were much higher for female applicants biased against associating women with professionalism, compared with female applicants with less bias (the coefficient on the bias x treatment term is 0.26 with a standard error of 0.11). Taken all together, these results suggest that the provision of the training program may have helped women overcome biases that would have otherwise held them back from employment in this sector, which is consistent with other research that has found stereotypes to be malleable to external influences, such as exposure to female politicians (Beaman et al. 2009). 10 The results in other columns of Table 7 also suggest that there were no statistically significant heterogeneous treatment effects for the other implicit biases we measured at baseline – namely, associating women with the home rather with the office and associating women with petty trade rather than an office job. For female applicants, the treatment interaction with the home-office bias is close to zero (0.006), as is the interaction with the petty trade-office bias (0.008). It is possible that these biases are more deep-seated and less malleable than the bias related to women’s professionalism. It is also possible that these biases do not affect the behavior of this select population that has attended university and indicated an interest in office work by applying to the program. Gender bias of any sort does not alter treatment impacts for males. Table 8 provides some suggestive evidence that the training experience in the ACCESS Nigeria program could have changed how women viewed their own abilities relative to men’s. It lists the subject specific scores on the final assessment. While their total scores were statistically indistinguishable from men’s, women did score significantly higher on precisely the oral skills that would have been observable to peers (voice clarity, fluency and vocabulary, grammar, accent, message clarity, language, grammar, and sentence construction), as opposed to those for which trainees would have been tested via computer (such as keyboard skills, internet and browsing skills, MS Office tools, numerical ability, analytical and logical ability). While these are only 10 While there is a literature examining whether changes in IAT measures reflect malleability of attitudes or malleability of the IAT measure itself (for example, Han et al., 2010), this paper is only inferring a change in attitudes from a group’s greater employment response to the treatment. The end line survey did not include a second administration of the IAT. 19 suggestive patterns, they suggest a mechanism through which training could have shifted beliefs for a subgroup of women. Robustness One challenge to the robustness of results in our sample is attrition. While 91 percent of the sample was successfully reached for end line interviews, there is a correlation between treatment status and attrition: 92 percent of treatment respondents responded to our survey compared to 88 percent of control. In other words, 7.75 percent (142 out of 1,832) of the treatment respondents could not be reached for the end line survey, compared to 12.06 percent (143 out of 1,186) of the control group. 11 To address this issue, we simulate alternate scenarios in which attrition rates were equivalent in treatment and control groups, under increasingly conservative assumptions about labor market outcomes for the attritors, in order to test the robustness of our results. 12 If 52 fewer respondents had attrited in control, attrition proportions would have been equal between treatment and control groups. Therefore, of the actual group of 143 control attritors, we randomly select 52 individuals and generate a simulated data set in which we assign outcome measures to them. (Adding these 52 individuals back into the sample makes the treatment and control attrition rates equivalent, with a treatment attrition of 7.75 percent (142/1832) and control attrition of 7.67 percent (91/1186)). We then simulate a world in which the 52 members of this “simulation” respondent group are more likely to get IT jobs compared to the true respondent control group rate of 5.8 percent. We estimate a range of scenarios, ranging from one where these new simulation respondents are 10 percent more likely than the rest of the control group to have gotten an ICT job, to scenarios where they are 200 percent more likely. We take 50 random draws of this group of 52 “simulation attritors” from the population of 143 total attritors, run the same regressions that we reported earlier for the regular sample, and report the p values on coefficients of interest including this simulated group. (We report average p values over the 50 simulations.) The results of this simulation are shown in Figure 6, which show that the treatment effect for the whole sample with and without strata controls loses significance at the 10 percent threshold if the 11 Other correlates of attrition include geographic variables and parental education. The results are robust to controlling for all of these variables. 12 One way that many researchers handle this kind of attrition is with Lee bounds. However, in our main specification, Lee bounds cannot be calculated due to cells without variation in attrition. 20 attritors reach 90 percent more likely to have obtained ICT jobs (represented by the value 1.9 on the x-axis of figure 1). However, the heterogeneous effects are highly robust, remaining significant for conventional levels of significance (p=0.05) even if the simulated non-attritors were three times more likely to get ICT jobs than the rest of the control sample (Figure 6, x-axis value of 3). Discussion and conclusion Encouraging skilled workers to shift to a new industry can be a challenge, particularly given labor market frictions such as a mismatch between the skills required by the industry and the skill endowment in the existing labor force, and gender norms that limit women’s mobility in the labor market (including implicit attitudes harbored by women themselves). We studied the role that these constraints may play in the labor market for recent university graduates in urban Nigeria through a randomized control trial that experimentally assigned slots in a job training program to prepare participants for work in the nascent ICT industry. This sector is of particular interest given that it has, in other settings, been less affected by social norms prohibiting females from accessing these relatively desirable, formal sector office jobs. In some countries, certain subsectors such as business-process outsourcing have even been female-dominated. Despite the training program’s coinciding with political violence and despite relatively low take- up of training offers, we find a modest average impact of the program. Those offered training were 1.7 percentage points or 26 percent more likely to be employed in the ICT sector two years later, with no significant impacts on overall employment. They were also almost 8 percent more likely to be actively searching for work at the time of the end line survey. Heterogeneous treatment effects can shed light on the extent to which skills, location, and gender norms moderate these effects. Treatment impact on ICT employment appears completely driven by applicants who had scored above the median on an initial assessment modeled after the new certification examination for the industry; these participants were 4.8 percentage points (75 percent) more likely to work in the ICT sector. This population was already more likely to work in the sector, and the program augmented this skill premium in employment, suggesting relatively high returns to industry-relevant skills. Impacts also varied across cities suggesting a mismatch between the supply of ICT-relevant skills and firms’ demand for them. Some cities appear not to have absorbed program participants, even though after the training program, they experienced 21 sizable changes in the fraction of people who would qualify for entry-level work according to international standards. Other heterogenous treatment effects suggest that gender norms may significant hinder women’s mobility in the labor market. Women who were less likely to associate women with professionalism than men in an implicit association test showed treatment impacts that were two times higher than the impacts estimated for unbiased women. The program cost to deliver training to each individual was about $606, which places this program below the median in the set of job training programs reviewed in McKenzie (2017). We observe no impact on earnings and net employment two years after the training. Thus, on the face of it, this training falls into the group of skills programs that do not pass a simple cost-benefit test, as Blattman and Ralston (2015) argue. However, in this paper we document two important non- pecuniary benefits which resulted from the training. First, the program induced switching into the nascent ICT sector in Nigeria. Given the government’s focus on developing this sector and its identification of a skills gap as a major constraint to sectoral growth, this policy lever has proved somewhat effective in increasing the employment of people with relevant skills in ICT. The program effects were larger for individuals who had higher baseline skill levels, suggesting that, to the extent that labor markets are efficient, the aggregate stock of skills in this emerging sector should grow over time. More importantly, training programs in new sectors open up the possibility set for workers. On one level, this is evident from our switching results. The switching, however, is significantly more pronounced for women who hold implicit biases against women’s professionalism and it induced their movement into a sector that was male-dominated in Nigeria. This matters for the economy as a whole on two levels. First, occupational segregation is a major contributor worldwide to the gap between male and female earnings (World Bank 2011). Second, the lack of mobility across professions based on gender norms is a significant barrier to the efficient functioning of labor markets. This program contained no special gender focus. However, by expanding the possibility set, particularly for women who were less likely to take advantage of jobs in this sector, it contributed to increased efficiency in the labor market. If this or a similar 22 program were taken to a larger scale, the improved efficiency in labor allocation could potentially contribute to the growth of the ICT sector and the economy as a whole. 23 References African Economic Outlook, 2012. Accessed at: http://www.africaneconomicoutlook.org/theme/youth_employment/youth-in-african-labour- markets/who-are-the-unemployed-discouraged-inactive-youth-in-africa/ Alesina, Alberto, Paola Giuliano, and Nathan Nunn (2013), “On the Origin of Gender Roles: Women and the Plough,” Quarterly Journal of Economics, 128(2): 469-530. Al-Jazeera. “Nigeria’s Kano Under Curfew After Attacks.” http://www.aljazeera.com/news/africa/2012/01/2012120172742659579.html. Jan 21 2012. Andrews, Donald (1989), “Power in Econometric Applications,” Econometrica, 57(5): 1059-90. Beaman, Lori, Raghabendra Chattopadhyay, Esther Duflo, Rohini Pande, and Petia Topalova (2009), “Powerful Women: Does Exposure Reduce Prejudice,” Quarterly Journal of Economics, 124(4). Beaman, Lori, Esther Duflo, Rohini Pande, and Petia Topalova (2012), “Female Leadership Raises Aspirations and Educational Attainment for Girls: A Policy Experiment in India,” Science, 335(6068): 582-586. Behaghel, Luc Bruno Crepon, and Marc Gurgand. (2014). “Private and Public Provision of Counseling to Job-Seekers: Evidence from a Large Controlled Experiment.” American Economic Journal: Applied Economics, 6(4): 142-74 Behar, Alberto. “Does the Nigerian formal sector pay more than its informal sector?” CSAE Working Paper WPS/2013-21. Bertrand, Marianne, Dolly Chugh, and Sendhil Mullainathan (2005), “Implicit Discrimination,” American Economic Review Papers and Proceedings, 95(2): 94-98. Blattman, Chris and Ralston, Laura (2015). “Generating employment in poor and fragile states: Evidence from labor market and entrepreneurship programs.” Working paper. http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2622220 Blau, Francine and Lawrence Kahn (2016), “The Gender Wage Gap: Extent, Trends, and Explanations”, NBER Working Paper No. 21913. G. Bryan, S. Chowdhury and A. M. Mobarak (2014), “Under-Investment in a Profitable Technology: The Case of Seasonal Migration in Bangladesh,” Econometrica, 82(5): 1671-1748 Buser, Thomas, Muriel Niederle, and Hessel Oosterbeek (2014), “Gender, Competitiveness, and Career Choices,” Quarterly Journal of Economics, 129(3): 1409-1447. 24 Campos, Francisco; Goldstein, Markus; McGorman, Laura; Munoz Boudet, Ana Maria; Pimhidzai, Obert (2015), “Breaking the Metal Ceiling: Female Entrepreneurs Who Succeed in Male-Dominated Sectors, “ Policy Research Working Paper No. 7503. World Bank, Washington, DC. World Bank. https://openknowledge.worldbank.org/handle/10986/23469 Eduquity (2012), “Access Nigeria: Post-training Assessment Report: Overall Performance, Demographic, and Comparative Analysis,” Eduquity Career Technologies Private Ltd. Escriche, Luisa (2007), "Persistence of Occupational Segregation: The Role of the Intergenerational Transmission of Preferences." The Economic Journal 117: 837-857. de Mel, Suresh, David McKenzie and Christopher Woodruff (2009), “Are Women More Credit Constrained? Experimental Evidence on Gender and Microenterprise Returns”, American Economic Journal: Applied Economics 1(3): 1-32 Filmer, Deon and Louise Fox (2014), Youth Employment in Sub-Saharan Africa. Africa Development Series. Washington, DC: World Bank. Fitzenberger, Bernd and Astrid Kunze (2005), "Vocational Training and Gender: Wages and Occupational Mobility Among Young Workers." Oxford Review of Economic Policy, Vol. 21, No.3: 392-415. Fortin, Nicole (2005), "Gender Role Attitudes and the Labour-market Outcomes of Women Across OECD Countries." Oxford Review of Economic Policy, vol. 21, No. 3: 416-438. Greenwald, Anthony G., Debbie E. McGhee, and Jordan L.K.Schwartz (1998), “Measuring Individual Differences in Implicit Cognition: The Implicit Association Test,” Journal of Personality and Social Psychology, 74(6): 1464-1480. Greenwald, A. G., Poehlman, T. A., Uhlmann, E. L., and Banaji, M. R. (2009). Understanding and using the Implicit Association Test: III. Meta-analysis of predictive validity. Journal of personality and social psychology, 97(1):17.Han, H. Anna, Sandor Czellar, Michael A. Olson, and Russel H. Fazio (2010), “Malleability of Attitudes or Malleability of the IAT,” Journal of Experimental Social Psychology, 46(2): 286-298. Jensen, Robert (2012), “Do Labor Market Opportunities Affect Young Women’s Work and Family Decisions? Experimental Evidence from India”, Quarterly Journal of Economics 127: 753-792 Lowes, Sara, Nathan Nunn, James A. Robinson, and Jonathan Weigel (2015), “Understanding Ethnic Identity in Africa: Evidence from the Implicit Association Test (IAT),” American Economic Review Papers and Proceedings, 105(5): 340-345. 25 Kiefer, Amy K. and Denise Sekaquaptewa (2007), “Implicit stereotypes and women’s math performance: How implicit gender-math stereotypes influence women’s susceptibility to stereotype threat,” Journal of Experimental Social Psychology, 43(5): 825-832. National Bureau of Statistics (2010), National Manpower Stock and Employment Generation Survey, Nigeria. Nopo, Hugo, Miguel Robles, and Jaime Saavedra (2008), "Occupational training to reduce gender segregation: The impacts of ProJoven," Economía Vol. XXXI, No. 62: 33-54. Nosek, Brian A., Mahzarin Banaji, and Anthony G. Greenwald (2002), “Harvesting Implicit Group Attitudes and Beliefs from a Demonstration Web Site,” Group Dynamics: Theory, Research, and Practice, 6(1): 101- 115. Nosek, Brian A., Frederick L. Smyth, N. Sriram, Nicole M. Lindner, Thierry Devos, Alfonso Ayala, Yoav Bar-Anan, Robin Bergh, Huajian Cai, Karen Gonsalkorale, Selin Kesebir, Norbert Maliszewski, Félix Neto, Eero Olli, Jaihyun Park, Konrad Schnabel, Kimihiro Shiomura, Bogdan Tudor Tulbure, Reinout W. Wiers, Mónika Somogyi, Nazar Akrami, Bo Ekehammar, Michelangelo Vianello, Mahzarin R. Banaji and Anthony G. Greenwald (2009), “National differences in gender-science stereotypes predict national sex differences in science and math achievement,” Proceedings of the National Academy of Sciences, 106(26): 10593-10597. Spencer, Steven J., Claude M. Steele, and Diane M. Quinn (1999), “Stereotype Threat and Women’s Math Performance,” Journal of Experimental Social Psychology, 35: 4-28. World Bank (2011), World Development Report 2012: Gender Equality and Development, World Bank Group, Washington D.C. World Bank (2012), “Realizing Nigeria’s Potential for the IT Enabled Services: Assessment of Core Competencies for Employability in the IT Enabled Service Sectors,” Policy Note, World Bank Group, Washington DC. World Bank (2014), World Development Report 2015: Mind, Society, and Behavior, World Bank Group, Washington D.C. 26 Figure 1: Timeline of program and study activities Jan - Mar 2011 Mar - Apr 2011 Nov - Dec 2011 Feb - April 2012 Apr 2012 Dec 2012 Feb - Apr 2014 Assessment Web-based and baseline Applicants Assessment employment Endline Advertising and outreach of applicants informed if 10 weeks of and network and phone for program by computer they have slot training certification job fair for survey of all in 5 training in program of trainees all applicants centers applicants 27 Figure 2: Experimental design, compliance, and attrition Assessed at baseline (n=3,092) Excluded from randomization Did not complete competency assessment (n=74) Excluded from Implicit Association Test Left test early (n=1,092) Randomized (n=3,018) Control applicants (n=1,186) Treatment applicants (n=1,832) Received training (n=0) Received training (n=1,007) Did not receive training (n=1,186) Did not receive training (n=825) Lost to follow-up (n=104) Lost to follow-up (n=142) 28 Figure 3: Explicit and implicit biases related to gender and professionalism 1 0.8 0.6 0.4 Female 0.2 Male 0 -0.2 Explicit Implicit Notes: Explicit and implicit scores are not directly comparable. Explicit associations were measured by questions that asked respondents to rate the professionalism of university educated men and women on a scale of 0 to 10. This graph presents the average of the within-respondent normalized difference between men and women's ratings. Implicit associations were measured by Implicit Association Tests. This graph presents the D-score (Greenwald et al, 1998), or the within-participant standardized difference between the sorting times for the different pairings of groups (men and women) with concepts (professional and unprofessional). For both explicit and implicit ratings, higher scores indicate an easier association between men and professionalism. D-scores between 0 and ±0.15 are considered to indicate little to no bias. Absolute deviations between 0.15 and 0.35, 0.35 to 0.65, and greater than 0.65 are considered to be slight, moderate, and strong associations, respectively. Figure 4: Implicit biases related to gender and career 0.6 0.5 0.4 0.3 0.2 0.1 0 Female Male Notes: Implicit associations were measured by Implicit Association Tests. This graph presents the D-score (Greenwald et al, 1998), or the within-participant standardized difference between the sorting times for the different pairings of groups (men and women) with the concepts home and office. Higher scores indicate an easier association between men and a career. Scores between 0 and ±0.15 are considered to indicate little to no bias. Absolute deviations between 0.15 and 0.35, 0.35 to 0.65, and greater than 0.65 are considered to be slight, moderate, and strong associations, respectively. Figure 5: Implicit biases related to gender and sector 0.5 0.4 0.3 0.2 0.1 0 Female Male Notes: Implicit associations were measured by Implicit Association Tests. This graph presents the D-score (Greenwald et al, 1998), or the within-participant standardized difference between the sorting times for the different pairings of groups (men and women) with the concepts office and petty trade. Higher scores indicate an easier association between men and an office career. Scores between 0 and ±0.15 are considered to indicate little to no bias. Absolute deviations between 0.15 and 0.35, 0.35 to 0.65, and greater than 0.65 are considered to be slight, moderate, and strong associations, respectively. Figure 6: Attrition simulation p-values Table 1: Descriptive statistics and treatment balance at baseline (1) (2) (3) (4) (5) (6) Descriptive statistics Balance at baseline p-value Treatment Comparison Mean SD N of mean mean difference Stratifying variables Female 0.36 0.48 3018 0.36 0.35 0.935 National Youth Service Corp 0.35 0.48 3018 0.35 0.35 0.952 Final year student 0.46 0.50 3018 0.46 0.47 0.787 Above median assessment score 0.50 0.50 3018 0.50 0.50 0.811 Abuja state 0.12 0.32 3018 0.12 0.12 0.846 Kano state 0.30 0.46 3018 0.30 0.30 0.900 Kaduna state 0.20 0.40 3018 0.20 0.20 0.894 Enugu state 0.20 0.40 3018 0.20 0.20 0.960 Lagos state 0.18 0.39 3018 0.18 0.18 0.943 Other baseline variables Age 25.96 2.91 3018 26.01 25.88 0.257 Christian 0.71 0.45 3018 0.72 0.70 0.208 Any labor market experience 0.82 0.39 3017 0.82 0.81 0.429 Currently employed 0.17 0.37 3018 0.16 0.18 0.170 Self-employed 0.11 0.31 3018 0.11 0.11 0.989 Any IT training 0.75 0.43 3018 0.76 0.74 0.219 Household size 7.55 5.57 2991 7.48 7.67 0.348 Mother has university education 0.25 0.43 2921 0.24 0.26 0.296 Father has university education 0.29 0.45 2791 0.29 0.29 0.677 Lives with parents 0.43 0.49 3018 0.43 0.42 0.720 Expenditures last 30 days (naira) 15619.50 22668.51 2984 15488.17 15821.96 0.694 Financed school with scholarship 0.10 0.30 3018 0.10 0.11 0.463 Aspires to work in IT industry 0.19 0.39 3018 0.20 0.19 0.489 Took IAT module 0.64 0.48 3018 0.65 0.63 0.283 IAT score: Home vs. office 0.49 0.47 1912 0.49 0.49 0.949 IAT score: Professional vs. unprofessional 0.07 0.39 1918 0.08 0.06 0.117 IAT score: Office vs. small business 0.37 0.40 1898 0.36 0.39 0.145 Notes: States refer to the state in which program applicants took the baseline assessment, survey, and IAT module. Table 2: Compliance and baseline scores by strata (1) (2) Dependent variable: Treatment: Full sample: Attended training Baseline assessment score Female -0.024 -0.068*** (0.024) (0.025) National Youth Service Corps (baseline) -0.096*** 0.096** (0.036) (0.039) Final year student (baseline) 0.028 -0.014 (0.032) (0.034) Above-median assessment score (baseline) 0.050** (0.024) Kano state -0.026 -0.26*** (0.040) (0.042) Kaduna state -0.062 -0.43*** (0.043) (0.046) Lagos state -0.13*** 0.22*** (0.045) (0.047) Enugu state -0.38*** 0.075 (0.044) (0.047) Number of observations 1,832 3,018 Mean 0.55 3.27 Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions contain a constant term, not reported here. Abuja state is the omitted geographic category. The standard deviation of the baseline assessment score was 0.695. Table 3: Fraction of the treatment group who accepted treatment meeting an international benchmark score (1) (2) (3) Percent meeting international benchmark N Before training After training Total 0.16 0.49 1,007 (0.01) (0.02) Abuja 0.24 0.58 120 (0.04) (0.05) Lagos 0.26 0.65 159 (0.03) (0.04) Enugu 0.21 0.61 198 (0.03) (0.03) Kano 0.12 0.40 302 (0.02) (0.03) Kaduna 0.07 0.33 228 (0.02) (0.03) Males 0.16 0.47 657 (0.01) (0.02) Females 0.16 0.52 350 (0.02) (0.03) Notes: This table is restricted to members of the treatment group who took up the offer of training and therefore took the assessment/certification exam at the end of training. International BPO firms use a benchmark score of 4 for entry-level positions. Table 4: Employment impacts of training program - ITT effects (1) (2) (3) (4) (5) (6) Employed Hours ICT job Any job Ln in ICT Employed worked search search earnings sector Treatment 0.017* -0.0080 0.73 0.00024 0.054*** 0.18 (0.0100) (0.018) (1.24) (0.012) (0.018) (0.21) - -0.060*** -0.10*** -9.49*** -0.0082 -1.00*** Female 0.042*** (0.0094) (0.019) (1.20) (0.011) (0.018) (0.21) National Youth Service Corps (baseline) 0.013 0.053* 2.82 -0.020 0.088*** -0.024 (0.016) (0.028) (1.97) (0.019) (0.029) (0.32) Final year student (baseline) -0.010 -0.0098 -1.07 -0.036** 0.080*** -0.21 (0.013) (0.025) (1.73) (0.017) (0.025) (0.27) Above-median assessment score 0.058*** 0.083*** 2.63** 0.068*** -0.028 0.62*** (baseline) (0.010) (0.019) (1.26) (0.012) (0.018) (0.21) Kano state 0.025 -0.091*** -5.37** 0.015 -0.0013 -0.48 (0.017) (0.030) (2.18) (0.019) (0.032) (0.36) Kaduna state -0.0042 -0.11*** -7.85*** 0.0039 0.044 -0.92** (0.017) (0.033) (2.35) (0.020) (0.034) (0.38) Lagos state 0.045** -0.048 -1.32 0.070*** 0.051 0.0070 (0.020) (0.034) (2.47) (0.024) (0.035) (0.40) Enugu state 0.0054 -0.22*** -9.41*** 0.022 0.032 -1.48*** (0.018) (0.034) (2.42) (0.022) (0.035) (0.40) Number of observations 2,709 2,733 2,673 2,731 2,733 2,694 Comparison group mean 0.06 0.64 34.37 0.10 0.71 6.45 Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions contain a constant term, not reported here. A total of 2,733 respondents were interviewed at endline. Sufficient information to classify the sector of respondents and calculate hours worked and earnings could only be obtained for subsets of these respondents. Table 5: Employment impacts of training program -TOT effects (1) (2) (3) (4) (5) (6) ICT Employed Hours Any job Ln Employed job in ICT worked search earnings search PANEL A: Participation from self-reports Training 0.030* -0.014 1.30 0.00042 0.096*** 0.32 (0.018) (0.033) (2.21) (0.021) (0.032) (0.37) Number of observations 2,733 2,733 2,673 2,731 2,733 2,694 PANEL B: Participation from administrative records Training 0.031* -0.015 1.33 0.00043 0.099*** 0.33 (0.018) (0.034) (2.26) (0.022) (0.033) (0.38) Number of observations 2,733 2,733 2,673 2,731 2,733 2,694 Note: Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions contain a constant term, not reported here and indicators for all stratifying variables used in the randomized assignment to treatment: gender, being above the median assessment score, state where assessment taken, being a final year student, and participation in National Youth Service Corps. Table 6: Heterogeneous impacts of training program: ITT effects (1) (2) (3) (4) ICT employment ICT search Employment Any search Panel A: Gender Treatment 0.018 -0.017 -0.025 0.052** (0.014) (0.016) (0.023) (0.022) Female -0.059*** -0.073*** -0.13*** -0.012 (0.014) (0.018) (0.031) (0.031) Treatment x female -0.0019 0.050** 0.049 0.0068 (0.018) (0.023) (0.039) (0.038) Panel B: Baseline assessment score Treatment -0.0064 -0.018 -0.0063 0.035 (0.011) (0.014) (0.027) (0.025) Above-median score (baseline) 0.029* 0.045** 0.085*** -0.052* (0.016) (0.019) (0.029) (0.030) Treatment x above-median score 0.048** 0.037 -0.0035 0.039 (0.020) (0.024) (0.037) (0.036) Panel C: Original testing site Treatment 0.077*** 0.051* -0.012 0.094* (0.024) (0.031) (0.049) (0.055) Kano state 0.064*** 0.070** -0.12** 0.041 (0.020) (0.028) (0.048) (0.052) Kaduna state 0.058*** 0.047 -0.087* 0.051 (0.021) (0.029) (0.051) (0.055) Lagos state 0.079*** 0.10*** -0.065 0.047 (0.025) (0.034) (0.052) (0.057) Enugu state 0.036* 0.027 -0.21*** 0.088 (0.020) (0.029) (0.053) (0.055) Treatment x Kano -0.063** -0.089** 0.039 -0.069 (0.031) (0.037) (0.060) (0.064) Treatment x Kaduna -0.100*** -0.069* -0.045 -0.012 (0.031) (0.038) (0.064) (0.067) Treatment x Lagos -0.056 -0.050 0.028 0.0071 (0.037) (0.046) (0.065) (0.069) Treatment x Enugu -0.050 -0.0090 -0.014 -0.090 (0.031) (0.040) (0.065) (0.068) Panel D: Intership status Treatment 0.015 -0.0035 -0.021 0.057** (0.012) (0.015) (0.023) (0.022) National Youth Service Corp 0.010 -0.027 0.030 0.093** (0.020) (0.024) (0.037) (0.038) Treatment x NYSC 0.0054 0.011 0.038 -0.0077 (0.022) (0.025) (0.038) (0.038) Notes : Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. All regressions contain the following, which are not reported here for to conserve space: a constant term and indicators for all stratifying variables used in the randomized assignment to treatment: gender, being above the median assessment score, state where assessment taken, being a final year student, and participation in National Youth Service Corps. Abuja is the omitted geographical location. Table 7: Implicit bias and heterogeneous treatment impacts on ICT employment (1) (2) (3) (4) (5) (6) (7) (8) (9) Females & Females Males males IAT Explicit Home vs. Office vs. Explicit Home vs. Office vs. Professionalism Professionalism sample bias office petty trade bias office petty trade Treatment 0.028** 0.022 0.029** 0.016 0.013 0.042** 0.034 0.059** 0.054* (0.013) (0.014) (0.014) (0.021) (0.015) (0.021) (0.023) (0.024) (0.029) Gender-professionalism 0.0013 -0.023 0.023** -0.010 (0.0037) (0.015) (0.011) (0.046) Treat x Gender-professionalism -0.0072 0.078** -0.0084 -0.0070 (0.0080) (0.036) (0.015) (0.057) Home vs. office 0.0094 0.073** (0.023) (0.030) Treat x Home vs. office 0.0067 -0.058 (0.034) (0.040) Small business vs. office 0.0086 0.063 (0.010) (0.042) Treat x Small business vs. office 0.026 -0.046 (0.029) (0.051) Number of observations 1,735 682 678 677 672 1,049 1,046 1,041 1,034 Comparison group mean 0.07 0.02 0.02 0.02 0.02 0.10 0.10 0.10 0.10 Notes: Note: Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions contain a constant term, not reported here and indicators for all stratifying variables used in the randomized assignment to treatment: gender (when specification is not restricted to a particular gender), being above the median assessment score, state where assessment taken, being a final year student, and participation in National Youth Service Corps. Biases are measured with D- scores. Sample sizes vary across different biases measured because of the way Implicit Association Tests are scored. Data is considered missing for respondents who make too many classification errors or take too long to respond to an item. Table 8: Final assessment scores by subject and gender (1) (2) (3) Females Males p-value Voice clarity 3.73 3.51 0.001 Fluency and vocabulary 3.70 3.54 0.021 Grammar 3.55 3.42 0.043 Accent 3.35 3.21 0.016 Message clarity 4.00 3.81 0.006 Language, grammar, and sentence construction 5.10 4.88 0.000 Listsening comprehension 4.34 4.42 0.286 Reading comprehension 5.27 5.12 0.036 Keyboard skills 3.49 3.50 0.892 Internet and browsing skills 3.24 3.35 0.171 MS Office tools 2.78 2.95 0.045 Numerical ability 3.52 3.61 0.193 Analytical and logical ability 4.08 4.10 0.817 Attention to detail 4.79 4.67 0.116 Total score 3.92 3.86 0.153 Number of observations 350 657 Appendix Table 1: Example test structure for gender and professionalism Implicit Association Test Block Number of trials Items assigned to left-key response Items assigned to right-key response 1 16 Faces of males Faces of females 2 16 Professional words Unprofessional words 3 16 Professional words + faces of males Unprofessional words + faces of females 4 32 Professional words + faces of males Unprofessional words + faces of females 5 32 Faces of females Faces of males 6 16 Professional words + faces of females Unprofessional words + faces of males 7 32 Professional words + faces of females Unprofessional words + faces of males Note: This table presents one of two sequences of blocks used for the gender-professionalism IAT. In another sequence, respondents were first asked to pair professional words with female faces and unprofessional words with male faces. Respondents were randomly assigned to a sequence. Blocks 3, 4, 6, and 7 are used to compute the d-score, the main measure of implicit association used in the analysis. Appendix Table 2: Stimuli for Implicit Association Test module Bias Stimuli A professional has high expertise and is honest, responsible, and hardworking. Those who have low expertise and are dishonest, careless, or lazy are unprofessional . How professional are younger females with a university degree (10=very Explicit: Gender and professionalism professional, 0=very unprofessional)? How professional are younger males with a university degree (10=very professional, 0=very unprofessional)? Terms for professional: Honest, competent, hardworking, responsible Implicit: Gender and professionalism Terms for unprofessional: Lazy, incompetent, dishonest, careless Terms for office: Conference, executive, manager, salary Implict: Office and home Terms for home: Family, marriage, kitchen, children Terms for office: Conference, executive, manager, salary Implicit: Office and petty trade Terms for petty trade: Trading, secondhand, market, kekenapep Appendix Table 3: Attrition (1) (2) (3) (4) (5) (6) (7) (8) (9) Full Treatment Control Interviewed p value of Interviewed p value of Interviewed p value of Attritor Attritor Attritor at endline difference at endline difference at endline difference Female 0.35 0.39 0.26 0.35 0.39 0.33 0.35 0.38 0.55 Age 25.98 25.70 0.12 26.02 25.79 0.36 25.92 25.62 0.24 Christian 0.71 0.71 0.93 0.72 0.72 0.98 0.70 0.70 0.97 National Youth Service Corp 0.34 0.42 0.01 0.34 0.42 0.06 0.34 0.41 0.09 Final year student 0.46 0.45 0.62 0.47 0.42 0.25 0.46 0.48 0.68 Any labor market experience 0.82 0.83 0.53 0.82 0.83 0.78 0.81 0.83 0.49 Currently employed 0.17 0.17 0.79 0.16 0.14 0.54 0.17 0.20 0.41 Self-employed 0.11 0.11 0.81 0.11 0.12 0.64 0.11 0.10 0.90 Any IT training 0.75 0.79 0.14 0.76 0.82 0.07 0.74 0.76 0.68 Total score on assessment 3.26 3.37 0.01 3.27 3.39 0.05 3.24 3.35 0.08 Above median assessment score 0.49 0.55 0.09 0.50 0.54 0.31 0.49 0.55 0.15 Abuja state 0.12 0.15 0.08 0.11 0.18 0.03 0.12 0.13 0.73 Kano state 0.29 0.39 0.00 0.29 0.37 0.04 0.29 0.40 0.01 Kaduna state 0.21 0.12 0.00 0.21 0.09 0.00 0.21 0.15 0.09 Enugu state 0.21 0.15 0.02 0.21 0.12 0.01 0.20 0.17 0.41 Lagos state 0.18 0.20 0.50 0.18 0.24 0.07 0.19 0.15 0.36 Household size 7.57 7.44 0.73 7.48 7.50 0.95 7.71 7.39 0.51 Mother has university education 0.24 0.31 0.02 0.23 0.34 0.01 0.26 0.27 0.72 Father has university education 0.28 0.36 0.01 0.28 0.39 0.01 0.28 0.33 0.18 Lives with parents 0.42 0.44 0.57 0.42 0.48 0.21 0.42 0.41 0.66 Financed school with scholarship 0.10 0.12 0.48 0.10 0.09 0.71 0.10 0.14 0.20 Aspires to work in IT industry 0.19 0.21 0.33 0.19 0.22 0.50 0.18 0.21 0.44 Took IAT module 0.63 0.69 0.08 0.64 0.72 0.07 0.62 0.66 0.44 Notes: States refer to the state in which program applicants took the baseline assessment, survey, and IAT module. Appendix Table 4: Participation in Implicit Association Test Took IAT Treatment -0.017 (0.016) Female -0.077*** (0.017) National Youth Service Corps -0.032 (0.025) Final year student -0.10*** (0.022) Above median assessment score 0.0037 (0.017) Kano state 0.12*** (0.026) Kaduna state 0.49*** (0.030) Lagos state -0.031 (0.028) Enugu state 0.35*** (0.032) Christian -0.0042 (0.020) Age -0.0017 (0.0031) Any labor market experience -0.021 (0.022) Wealth index 0.0097* (0.0052) Any IT training -0.017 (0.019) Employed 0.00076 (0.022) Self-employed -0.042 (0.026) Took endline -0.0018 (0.027) Number of observations 3,012 Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions contain a constant term, not reported here. Abuja state is the omitted geographic category. 41 Appendix Table 5: Employment impacts of training program - ITT effects with controls (1) (2) (3) (4) (5) (6) Employed Hours ICT job Any job Ln Employed in ICT worked search search earnings Treatment 0.017* -0.011 0.39 0.010 0.050** 0.16 (0.0099) (0.018) (1.23) (0.013) (0.020) (0.21) Female -0.058*** -0.077*** -8.29*** -0.036*** 0.0033 -0.84*** (0.0097) (0.019) (1.23) (0.012) (0.020) (0.22) National Youth Service Corps 0.012 0.045 2.48 -0.023 0.055* -0.044 (0.016) (0.028) (1.97) (0.021) (0.031) (0.32) Final year student -0.0077 0.020 0.80 -0.043** 0.066** 0.022 (0.014) (0.025) (1.72) (0.019) (0.027) (0.28) Above median assessment 0.057*** 0.11*** 3.55*** 0.063*** -0.032 0.79*** score (0.011) (0.019) (1.27) (0.013) (0.020) (0.21) Kano state 0.030* -0.070** -4.14* 0.0071 0.014 -0.34 (0.017) (0.030) (2.15) (0.021) (0.035) (0.36) Kaduna state 0.0079 -0.11*** -6.67*** 0.019 0.065* -0.90** (0.017) (0.033) (2.39) (0.021) (0.037) (0.39) Lagos state 0.051** -0.045 -1.27 0.067*** 0.051 0.068 (0.020) (0.033) (2.44) (0.026) (0.038) (0.40) Enugu state 0.0090 -0.16*** -6.85*** 0.022 0.022 -0.95** (0.019) (0.035) (2.46) (0.025) (0.038) (0.41) Age 0.0012 0.026*** 1.23*** -0.0018 0.0026 0.16*** (0.0019) (0.0032) (0.22) (0.0023) (0.0035) (0.037) Ln Expenditures -0.00018 -0.0047 -0.27 0.0024 -0.0029 -0.0013 (0.0030) (0.0057) (0.39) (0.0040) (0.0073) (0.062) Christian 0.020* -0.023 1.47 0.021 0.040* -0.31 (0.012) (0.023) (1.56) (0.014) (0.024) (0.25) Any experience at baseline -0.0073 0.065*** 2.36 -0.0054 0.038 0.44 (0.012) (0.025) (1.61) (0.016) (0.027) (0.27) Employed at baseline 0.011 0.065*** 3.29* 0.017 -0.032 0.79*** (0.014) (0.024) (1.71) (0.018) (0.027) (0.27) Self-employed at baseline 0.037** 0.045 6.91*** -0.0033 -0.038 0.57* (0.018) (0.028) (2.13) (0.020) (0.031) (0.32) Any IT training at baseline 0.037*** 0.012 2.45* 0.037*** 0.021 0.076 (0.0098) (0.021) (1.40) (0.013) (0.023) (0.24) Number of observations 2,728 2,728 2,668 2,331 2,333 2,689 Comparison group mean 0.064 0.644 34.369 0.103 0.707 6.445 Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions contain a constant term, not reported here. A total of 2,733 respondents were interviewed at endline. Sufficient information to classify the sector of respondents and calculate hours worked and earnings could only be obtained for subsets of these respondents. 42 Appendix Table 6: ANCOVA estimates - ITT effects (2) (3) (4) (6) Formal sector Self- Employed Ln earnings employment employment Treatment -0.0067 -0.024 0.017 0.056 (0.018) (0.018) (0.015) (0.039) Female -0.11*** 0.0031 -0.062*** -0.15*** (0.019) (0.018) (0.015) (0.040) National Youth Service Corps 0.050* 0.070** -0.021 0.060 (0.028) (0.029) (0.025) (0.062) Final year student -0.0044 -0.012 -0.026 -0.056 (0.025) (0.023) (0.021) (0.053) Above median assessment score 0.082*** 0.14*** -0.040** 0.12*** (0.019) (0.018) (0.016) (0.039) Kano state -0.089*** -0.061* -0.042 -0.27*** (0.030) (0.033) (0.028) (0.066) Kaduna state -0.11*** -0.10*** -0.054* -0.39*** (0.033) (0.034) (0.030) (0.069) Lagos state -0.039 -0.055 -0.010 -0.23*** (0.034) (0.037) (0.031) (0.071) Enugu state -0.21*** -0.16*** -0.071** -0.55*** (0.035) (0.034) (0.030) (0.071) Employed at baseline 0.086*** (0.024) Employed in formal sector at -0.035 baseline (0.023) Self-employed at baseline 0.070*** (0.027) Ln expenditures at baseline 0.012*** (0.0031) Number of observations 2,733 2,733 2,733 2,714 Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions contain a constant term, not reported here. A total of 2,733 respondents were interviewed at endline. Sufficient information to classify the sector of respondents and calculate hours worked and earnings could only be obtained for subsets of these respondents. 43