The World Bank Economic Review, 39(1), 2025, 143–163 https://doi.org10.1093/wber/lhae019 Article Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 The Gendered Impact of Digital Jobs Platforms: Experimental Evidence from Mozambique Sam Jones and Kunal Sen This study examines the impact of digital labor-market platforms on jobs outcomes using a randomized en- couragement design embedded in a longitudinal survey of Mozambican technical-vocational college graduates. We differentiate between platforms targeting formal jobs, where jobseekers direct their search, and informal tasks, where clients seek workers. Our analysis reveals statistically insignificant intent-to-treat and complier- average treatment effects for headline employment outcomes in the full sample. Notably, while the average male moderately benefits from platform usage, women do not. Rather, they are less responsive to the encouragement nudge, and female treatment compliers report higher reservation wages and lower job search. This suggests digital platforms can inadvertently perpetuate gender disparities in labor markets. JEL classification: J46, J64, J68, O15 Keywords: labor markets, digital platforms, gender, Mozambique 1. Introduction Youth employment is a major policy concern across Africa, with about one in four young persons not in employment, education, or training (Bandiera et al. 2022; ILO 2022a). Alongside structural challenges, one reason for this is labor-market frictions that impede good quality matches between employers and qualified candidates (Van Den Berg 1999; Chade, Eeckhout, and Smith 2017; Rud and Trapeznikova 2021). These encompass (high) search costs, such as basic transportation expenses (Mulalic, Van Om- meren, and Pilegaard 2014; Franklin 2018), screening costs (Abebe et al. 2021), inaccurate beliefs (Beam 2016; Abebe et al. 2021; Jones and Santos 2022), and other obstacles to labor-market information flows. Sam Jones (corresponding author) is a research fellow at UNU-WIDER, based at the Ministry of Economy and Finance, Mozambique; his email address is jones@wider.unu.edu. Kunal Sen is the Director of UNU-WIDER, Helsinki; his email address is sen@wider.unu.edu. Thanks to David McKenzie and anonymous reviewers for very constructive feedback. We are grateful to Ivan Manhique, Ricardo Santos, and Anna Schnupp for excellent research support, as well as to Giannis Panagiotou and Tiago Borges Coelho of UX (Emprego/Biscate) for great collaboration and data access. We also appreciate very helpful comments from Rute Caeiro, Gary Fields, Rob Garlick, Kalle Hirvonen, Rodrigo Oliveira, John Rand, and Miri Stryjan, as well as participants at the 2022 “Inclusive Growth in Mozambique” annual conference, the 2022 IZA/World Bank/NJD/UNU-WIDER “Jobs and Development” conference, and seminars at IFPRI and UNU-WIDER. The data used in this study were collected under the project “Inclusive Growth in Mozambique – Scaling-Up Research and Capacity” (Phases I and II), a collaboration between UNU-WIDER, University of Copenhagen, University of Eduardo Mondlane, and the Mozambican Ministry of Economy and Finance, financed through specific programme contributions by the governments of Denmark, Finland, Norway, and Switzerland. C UNU-WIDER 2024. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. This is an Open Access article distributed under the terms of the Creative Commons Attribution 3.0 IGO License (https://creativecommons.org/licenses/by/3.0/igo/) which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. 144 Jones and Sen Reflecting improved access to technology, including smartphones, digital jobs platforms have grown rapidly over recent years (ILO 2021). In theory, these can help lower the costs of job search, expand its geo- graphical scope, and alleviate informational frictions, making it more likely matches will occur (Carranza and Mckenzie 2024). In reality, however, digital platforms are subject to multiple challenges, including congestion and information asymmetries, which limit their effectiveness, particularly for new jobseek- Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 ers without a verifiable work history (Pallais 2014; Stanton and Thomas 2016). Echoing the “puzzle of ineffective internet job search” in advanced economies (Kroft and Pope 2014), a growing experimental literature in developing countries studies the labor-market effects of digital labor platforms but finds lim- ited positive impacts on employment or wages (Chakravorty et al. 2021; Kelley, Ksoll, and Magruder 2023). An exception is Wheeler et al. (2022), who train jobseekers to join and use an online networking platform in South Africa. They show the platform reduces information frictions and yields significant benefits, primarily at the extensive margin of employment (finding a first job). This paper provides causal evidence on the extent to which digital platforms help improve employment outcomes for recent college leavers in Mozambique. We add to the evidence base from a new context, a low-income African country with a large informal sector, and make two specific contributions to the literature. First, in contrast to most previous studies that focus on effects for the average participant, we investigate differential effects by gender. Understanding such heterogeneity is particularly relevant in low-income countries where employment outcomes and access to technology differ by a large margin between men and women. In Africa, young women are disproportionately observed in unpaid (domestic) or informal types of self-employment, with men dominating formal employment positions particularly in the private sector (ILO 2022b; Bandiera et al. 2022). As discussed by Jayachandran (2021), these gaps reflect gender norms associated with allocating responsibilities for household reproduction, as well as controls over female mobility and decision-making. Plausibly, lowering matching (search) frictions via digital labor platforms may be relatively more valu- able to women when they face greater difficulties in engaging in offline search or when they are more constrained in the jobs they can feasibly undertake (Chung and Van der Horst 2018). Yet a concern is that platforms merely reproduce or even accentuate the gendered segregation of labor markets. This can arise if barriers to using these platforms are higher for women, or if opportunities intermediated by these platforms are skewed toward activities where men hold existing advantages (Rodríguez-Modroño, Pesole, and López-Igual 2022; Adams-Prassl et al. 2023). For instance, platform or gig-work may be associated with greater safety risks for women as opposed to men, such as in the case of ride-sharing or delivery services (e.g., see Anwar 2022; Hamal and Huijsmans 2022; Rani et al. 2022). As a second contribution we provide a comparative evaluation of the impacts of two different plat- forms serving distinct labor-market segments. The first, Emprego, is for formal jobs, where jobseekers search and apply for opportunities posted by firms and organizations. The second, Biscate, seeks to match demand for specific tasks or services (e.g., plumbing, catering) to freelancers. This platform focuses on self-employed blue-collar workers active in the informal labor market and functions in a way that clients (primarily private individuals) search for workers in their location with relevant skills. The compari- son of different platforms is valuable since existing studies generally focus on conventional formal jobs boards (like Emprego), while platforms for informal task-based employment have received less attention in developing countries. Understanding the relative efficacy of different platforms has implications for platform design as well as for public policies to increase employment. In contexts of high informal self- employment and limited formal employment growth, there may be a role for public support to digital jobs platforms for informal task-based work, as long as evidence shows they genuinely improve outcomes for workers. Our focus is on graduates of Technical and Vocational Educational and Training (TVET) institutes in Mozambique. We embed an encouragement design into a school-to-work tracer survey, randomly allo- cating TVET graduates into one of three experimental arms: (a) to receive an SMS invitation to register The World Bank Economic Review 145 on Biscate; (b) to receive an SMS invitation to register on Emprego; and (c) a control group (no SMS). We estimate the impact of both platforms on a range of labor-market outcomes, covering extensive and inten- sive margins, including rates of employment, after-tax wage income, and job quality, as well as secondary job search outcomes such as reservation wages. Consistent with previous experimental studies (see Carranza and Mckenzie 2024), intent-to-treat (ITT) Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 treatment effects and instrumental variables estimates of the complier-average treatment effect (where compliance is defined in terms of platform registration) in our study are close to zero in the full sample for most headline labor-market outcomes for both platforms. In part, this reflects significant non-compliance, including moderate uptake in the treatment groups and some platform usage among controls. To address this, we implement an innovative approach that adjusts for observed non-compliance in our setting. The estimator we use relies on an assumption of principal ignorability, assigning greater weight to units whose baseline characteristics are more strongly associated with platform usage driven by the encouragement nudge (Ding and Lu 2017). This changes the estimand but allows more precise estimation, yielding indica- tive evidence of moderate positive average impacts for individuals successfully nudged to use the Biscate platform, but not Emprego. A paper nearest to our own is Afridi et al. (2022), who provide experimental evidence of the em- ployment effects of a “hyper-local” platform to recruit blue-collar workers in Delhi, focusing on how benefits from platform usage interact with social networks and gender roles. In a first treatment, they offer free registration to married couples, and in the second the same offer is provided to select members of the wife’s peer network. Their study finds no increase in the likelihood of women working in either treatment group one year after the intervention, but they find favorable labor-market outcomes for their husbands. Our analysis tends toward a sober yet nuanced assessment of the potential benefits of digital jobs platforms for young women. We find women are both less likely to use digital labor platforms and less responsive to our encouragement nudge. In turn, complier-average treatment effects for employment out- comes are broadly positive for the average man but not for the average woman. This result is obtained for both platforms but is somewhat larger on the formal jobs platform (Emprego), especially for metrics of job quality. Among female treatment compliers, we also find reservation wages are significantly higher than among men, but their search effort is comparatively lower (echoing Kelley, Ksoll, and Magruder 2023). This underscores a conclusion that digital platforms can have complex and gender-differentiated effects, which depend both on the labor-market segment they target and interactions with extant social norms. 2. Study Design 2.1. Background Our experiment was embedded in a longitudinal (tracer) survey of the school-to-work transitions of graduates of Technical and Vocational Educational and Training (TVET) institutes in Mozambique. As Jones, Santos, and Xirinda (2021) summarize, Mozambique began to reform its technical and professional education system in the early 2000s: in 2001, the government approved a new 10-year TVET strategy; in 2006 the World Bank launched a 15-year project to improve the quality, relevance, and responsiveness of the TVET system to the labor market; a new framework Vocational Education Law was passed in 2014, establishing a new regulatory authority; and in 2017, the National Professional Education Fund (Fundo Nacional de Educação Profissional) was established. So, at the end of almost two decades of reforms, the tracer survey sought to investigate how new TVET graduates fare in the labor market. The survey was undertaken in two phases. The first, which ran from October to November 2019, was an in-person baseline survey of final year students in TVET colleges selected to cover all regions of the country. Specifically, we visited institutions located in Maputo City, Maputo Province, Tete, Cabo Delgado, 146 Jones and Sen and Nampula provinces.1 The baseline survey collected information on students’ cognitive abilities and their family background, as well as expectations and aspirations for the future. The second phase, starting after the preceding academic year and running from January to November 2020, comprised four follow- up telephone surveys. These collected data on the evolution of labor-market outcomes of each participant over time. We sought to recontact each participant in each follow-up round, yielding a panel of four Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 quarterly observations per person (plus the baseline).2 2.2. Experiment Within the framework of the tracer survey we partnered with the operator of two locally developed digital labor-market platforms—Emprego and Biscate—which target distinct labor-market segments. The first, Emprego, meaning “job” or “employment” in Portuguese, is a conventional jobs board, where employers post vacancies and receive applications from registered users. The platform is only accessible via the inter- net (www.emprego.co.mz), including via a dedicated mobile application. On average, the site is accessed by around 18,000 individuals daily and over 1,400 organizations in Mozambique have used it, encom- passing private firms and non-governmental agencies. Effectively, the platform spans what in Mozambique is a comparatively small segment of the labor market, namely the market for formal jobs outside the pub- lic sector, of which most are professional services roles. Moreover, many vacancies posted on Emprego demand a comparatively high level of education (e.g. tertiary), previous professional experience, and/or English-language skills. Nonetheless, the site does include a smaller number of blue-collar vacancies, such as those with technical-vocational qualifications (see vacancy example fig. S1.1 in supplementary online appendix S1). Biscate, meaning “odd job,” is a platform to match demand and supply of informal freelancers for specific tasks or services. Unlike Emprego, the platform is set up to allow prospective clients to find local contractors offering specific services, such as plumbing or manicure, for direct payment in cash. The platform thus pertains to semi-skilled manual tasks demanded by private individuals outside the formal labor market (see fig. S1.2). Aside from agriculture, informal activities of this sort dominate the Mozambican economy but nonetheless can be a means to gain experience and contacts. The platform is accessed mainly via mobile phone on the Vodacom network using Unstructured Supplementary Service Data, which is not reliant on smartphone technology; it also has a dedicated smartphone application and website (www.biscate.co.mz). The platform has around 50,000 registered workers, and since its launch in 2016, more than 30,000 customers have used the service and 120,000 worker contacts have been requested through it. As noted, in addition to serving different labor-market segments, the two platforms differ in terms of the direction in which search takes place. Under Biscate, clients use the platform to contact potential workers based on their location and profile, as well as any comments or ratings from previous tasks.3 Under Emprego, candidates can post their CVs and contact potential employers, based on their presumed suitability for a specific job. Despite this difference, data from both platforms point to an excess supply of workers. Figures S2.1 and S2.2 (in the supplementary online appendix) show trends in the number of 1 Colleges were selected to yield a sample that is representative of the population of students attending large TVET institutes in each one of the selected provinces. As such, the survey covered 85 separate classes across 20 different colleges, with an average of 16 students responding per class. 2 Descriptive statistics from the survey, as well as further details on the sample structure, are found in Jones, Santos, and Xirinda (2021). Based on official information regarding the universe of TVET institutes, we construct post stratification weights to correct for minor disproportions in the share of sampled observations versus the regional distribution of final-year students. These are applied throughout but only imply minor adjustments. 3 This is relevant for later analysis of gender differences. Namely, gender-based preferences can be applied ex ante on Biscate, such as by only contacting men for a given task. Similarly under Biscate, individuals have to categorize themselves by profession, a feature that may accentuate gender divisions. The World Bank Economic Review 147 Figure 1. Count of Participants Classified by Experimental Status. Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 Source: Own estimates. Note: The flow chart summarizes and partitions the number of observations (N) in different survey rounds and groups; all lower nodes are subsets of higher nodes; ellipses refer to baseline survey; boxes refer to follow-up telephone rounds. contacts made by clients to prospective workers on Biscate from 2019 to early 2021 in our four main baseline survey locations. While a total of about 2,500 contact requests were made per month across these locations in 2020, the rate of contact, defined as the number of contacts per month per worker, was less than 0.2. For Emprego, around 70 new jobs were posted in each month of 2020 across the same four locations, of which only about 40 did not demand either higher education or more than three years of experience (fig. S2.3); but each of these jobs received an average of more than 300 applications (fig. S2.4).4 To test the contributions of the two platforms to employment outcomes we adopted an encouragement design. Before the start of the second round of the follow-up surveys we sent SMS invitations to individ- uals randomly selected from the baseline sample. The experimental intervention (nudge) comprised two separate treatment arms with no crossover, namely (a) an SMS invitation to register on Emprego and (b) an SMS invitation to register on Biscate. Allocation to these two treatment groups was stratified by gender and course type, with a moderately larger share of individuals with manual-type qualifications allocated to the Biscate nudge.5 Control group participants remained free to use any of the platforms, but received no encouragement (SMS). The target population for the intervention was the full baseline TVET sample (N = 1,639). However, to minimize contamination, particularly from individuals with prior experience of either of the platforms, we restricted the sample to an eligible subgroup. Concretely, from the full sample we excluded (a) indi- viduals who did not consent to participate in the follow-up telephone surveys, (b) individuals with shared or duplicated contact numbers in the baseline survey, (c) individuals without a Vodacom mobile phone contact number, and (d) individuals already registered on either the Biscate or Emprego platforms before the start of the first follow-up round. Illustrated in fig. 1, this yielded a subsample of 1,357 eligible par- ticipants, equal to 83 percent of the full baseline sample. The remainder of the analysis focuses on this group. 4 Further data from the platforms, such as pertaining to wage rates, are not available as these are generally not advertised and payments occur outside the platforms. 5 These stratifying variables are controlled for in all subsequent analysis. 148 Jones and Sen Table 1. Descriptive Statistics (I) (II) (III) (IV) (V) Round 1 Rounds 2–4 All Control Emprego Biscate Pr. Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 (a) Baseline covariates: Age 21.63 21.62 21.91 21.27 0.014 Female 0.44 0.38 0.46 0.40 0.986 Manual course 0.59 0.61 0.49 0.68 0.000 Public school 0.65 0.65 0.64 0.70 0.409 Work for self 0.80 0.77 0.81 0.81 0.234 Work for others 0.82 0.84 0.81 0.84 0.273 Prev. experience 0.45 0.48 0.44 0.43 0.103 Phone/computer/internet 0.69 0.70 0.71 0.67 0.674 Mother second. edu 0.53 0.49 0.58 0.53 0.148 Father second. edu 0.63 0.60 0.62 0.65 0.573 (b) Take-up: Emprego user 0.02 0.17 0.22 0.15 0.000 Biscate user 0.00 0.04 0.10 0.41 0.000 (c) Employment outcomes: Paid work 0.23 0.33 0.30 0.30 0.804 Labor income 38.18 47.68 39.87 40.10 0.207 Working 0.38 0.42 0.38 0.38 0.349 Hours worked (week) 18.01 18.15 16.51 17.26 0.788 Permanent position 0.05 0.06 0.07 0.06 0.934 Has a contract 0.04 0.08 0.09 0.07 0.397 Pays social security 0.07 0.10 0.10 0.10 0.942 Vertical match 0.37 0.36 0.33 0.35 0.786 Horizontal match 0.33 0.31 0.28 0.29 0.457 Employment score − 0.08 0.03 − 0.05 − 0.03 0.710 (d) Search behavior: Seeking work 0.80 0.69 0.70 0.72 0.064 Hours searching 9.14 4.48 4.43 4.96 0.164 Reservation wage 210.60 172.33 165.07 168.00 0.437 Obs. 1,357 2,376 1,188 1,299 – Source: Own estimates. Note: The cells report means for different survey rounds and subgroups; column (I) shows results for the full eligible sample in round 1, prior to the intervention; columns (II)–(IV) pool follow-up rounds 2 to 4, separating between control, Emprego, and Biscate treatment arms; panel (a) refers to fixed individual characteristics as captured at baseline; panel (b) gives average metrics of platform usage; panel (c) summarizes core employment outcomes; panel (d) are secondary metrics pertaining to job search; the final row gives the number of observations; the last column (Pr.) reports the probability that treatment group means jointly differ from those of the control where in panel (a) we only consider the pre-intervention period, but in the remaining panels we only consider the post-intervention period; incomes and reservation wages are reported in USD/month and take a zero value for individuals out of work (as do other employment outcomes). 2.3. Data and Descriptive Statistics Table 1 summarizes data from across the survey rounds, while table S2.1 disaggregates by relevant sub- samples (namely, gender and course type). Panel (a) reports individual information as collected at the baseline. Column (I) gives averages for the full eligible sample in round 1 of the phone survey, which is prior to the nudge intervention. Columns (II)–(IV) show means for different treatment groups, pooling the three rounds in the post-intervention period. Panel (b) reports data on usage of the two platforms. In the raw data we have three relevant can- didate metrics for platform uptake: (a) information reported to us by the platform on whether a given individual has registered, based on their phone number, (b) a self-reported measure of whether the in- dividual registered on the platform, and (c) a self-reported measure of whether the individual used the The World Bank Economic Review 149 platform to look for work, which is only relevant for Emprego where registered individuals search for jobs. These measures can differ for a variety of reasons—e.g., if the individual uses a different number to register, if their profile is incomplete, if they confuse registration with more basic usage, or if they browse the platforms without registering. Since the variables are all positively associated and in the ab- sence of a strong a priori view as to which is most informative, we calculate their simple average and use Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 this as a continuous measure of platform usage.6 As shown in the table, these confirm that individuals exposed to either of the two nudges reported a higher intensity of usage on the relevant platform than individuals in other arms—i.e., both nudges were partially effective in stimulating platform usage (see also table 2). Panel (c) reports the core set of employment outcomes, covering whether the individual has under- taken paid work in the past seven days, labor-market income, and various metrics of job quality such as whether they have a contract and the quality of the match to their education and training.7 All these outcomes remained comparatively low throughout the survey period. For instance, in the first round just 38 percent of the eligible group reported undertaking any work (paid or unpaid) in the seven days prior to being interviewed, increasing by just 1 percentage point in later rounds. Also, fewer than 10 percent of participants reported having a permanent job in any round. To facilitate later analysis and ameliorate multiple-hypothesis concerns, we combine all nine of these employment outcome variables into an overall score. Reported in the last row of panel (c), this is derived from the first principal component of the nine preceding variables, estimated from observations in the first follow-up round only. We regularize this synthetic score to take a mean of 0 and standard deviation of 1 in the control group. Panel (d) summarizes complementary information on job search behavior. Specifically, we include whether the individual reports to be actively seeking a(nother) job, the number of hours devoted to job search per week, and the self-reported reservation wage. The final column reports results from a joint test of equality between treatment and control groups. For panel (a) we focus on differences in round 1, akin to a balance test, and in remaining panels we focus on differences in the post-intervention period, akin to simple treatment effect regressions.8 In general, these results indicate no strong systematic associations between assignment to the interventions and baseline covariates or later outcomes. Some employment outcomes (such as being in paid work) appear marginally lower in the nudge (treatment) groups versus the control, but these differences are generally substantially smaller than the minimal detectable effect (of 0.0755) we estimated under simulations run prior to the experiment for our sample size. At the same time, there is indicative evidence of a slightly higher propensity to be looking for work in the treatment groups. Lastly, the final row reports the number of observations in each group. As further clarified in fig. 1, just five of the eligible sample were lost in the first follow-up round; and by the fourth round, more than a year after the baseline survey, we were able to contact 97 percent of the eligible sample (1,311 individuals), implying an extremely low rate of attrition. This is supported by fig. S2.5, which reports the sum of observations in each follow-up round by eventual experimental group status, confirming low attrition across all experimental arms.9 6 For Biscate, we take the mean of the self-reported and external measures of registration. Our main results are qualitatively unchanged if we take the maximum of the alternative measures. 7 All outcomes are set to zero for individuals not working or without wage income. Continuous outcomes are transformed using the inverse hyperbolic sine. 8 Test are based on separate OLS regressions in which (where relevant) we control for stratifying variables deployed in the randomization process and survey round fixed effects. These are individually excluded when they feature as dependent variables. 9 Due to very low attrition, we do not consider this as a material source of bias. 150 Jones and Sen Table 2. Effect of Encouragement Nudges on Digital Platform Usage (I) (II) (III) (IV) (V) (VI) (VII) (VIII) Platform usage → Emprego Biscate Any Ext. Self Srch Mean Ext. Self Mean Max Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 (a) Full sample: Emprego SMS 0.09∗∗∗ 0.10∗∗∗ 0.02∗ 0.08∗∗∗ 0.02∗∗ 0.13∗∗∗ 0.07∗∗∗ – (0.02) (0.02) (0.01) (0.01) (0.01) (0.02) (0.01) Biscate SMS − 0.01 0.01 − 0.01 − 0.00 0.47∗∗∗ 0.27∗∗∗ 0.37∗∗∗ – (0.01) (0.02) (0.01) (0.01) (0.03) (0.02) (0.02) Any SMS – – – – – – – 0.21∗∗∗ (0.02) R2 adj. 0.07 0.14 0.03 0.16 0.39 0.18 0.34 0.24 Control mean 0.03 0.18 0.07 0.11 0.01 0.06 0.03 0.14 (b) Male sample: Emprego SMS 0.13∗∗∗ 0.14∗∗∗ 0.04∗∗ 0.12∗∗∗ 0.01 0.14∗∗∗ 0.07∗∗∗ – (0.03) (0.03) (0.02) (0.02) (0.01) (0.03) (0.02) Biscate SMS − 0.01 0.03 − 0.01 0.01 0.52∗∗∗ 0.31∗∗∗ 0.41∗∗∗ – (0.01) (0.03) (0.02) (0.02) (0.03) (0.03) (0.03) Any SMS – – – – – – – 0.24∗∗∗ (0.02) R2 adj. 0.10 0.17 0.03 0.19 0.43 0.20 0.37 0.28 Control mean 0.04 0.21 0.09 0.14 0.02 0.08 0.05 0.17 (c) Female sample: Emprego SMS 0.04∗∗ 0.04 0.01 0.03∗ 0.01 0.10∗∗∗ 0.06∗∗∗ – (0.02) (0.03) (0.02) (0.02) (0.01) (0.02) (0.02) Biscate SMS − 0.00 − 0.02 0.00 − 0.01 0.40∗∗∗ 0.21∗∗∗ 0.30∗∗∗ – (0.01) (0.03) (0.02) (0.01) (0.04) (0.03) (0.03) Any SMS – – – – – – – 0.16∗∗∗ (0.02) R2 adj. 0.05 0.09 0.03 0.11 0.36 0.15 0.31 0.19 Control mean 0.02 0.14 0.05 0.09 0.00 0.03 0.01 0.09 Source: Own estimates. Note: The table summarizes results of estimates of equation (1) for different platforms (in columns) and measures of uptake (in subcolumns); panel (a) is the full sample, while panels (b) and (c) are separate estimates by gender; outcome “Ext.” takes a value of 1 if the individual has an externally verified profile on the platform; “Self” takes a value of 1 if the individual states they have a profile on the platform; “Srch” takes a value of 1 if the individual stated they used the platform to search for jobs (only relevant for Emprego); columns labeled “Mean” use the row-wise average of the three separate measures; selected regression coefficients shown; column (VIII) defines usage as the largest value of either the Emprego or Biscate mean; the final row of each panel reports sample average of the relevant dependent variable for the control group after round one; data cover all survey rounds and round fixed effects included throughout; standard errors (in parentheses) clustered by individuals. Significance: ∗ 10 percent, ∗∗ 5 percent, ∗∗∗ 1 percent. 3. Empirical Strategy We seek to estimate the causal impact of digital labor-market platforms on labor-market outcomes.10 A necessary condition for our experiment to identify the contribution of these platforms is that the SMS nudges boosted platform registration and subsequent usage. This represents the first testable hypothesis, reflected in the following general model: Usage j,it = α + β j [Nudge j = 1]it + Xit θ + ψt + it , (1) j 10 Our pre-analysis plan is set out in Jones and Santos (2020). Supplementary online appendix S3 comments on the con- sistency between the planned and final analysis. The World Bank Economic Review 151 where j indexes the focus platforms (Emprego or Biscate), i indexes individuals, and t is time. The main explanatory variables are the individuals’ experimental status, which take a value of 1 if they received an SMS nudge and 0 otherwise (being 0 in the first follow-up round for all); X is a vector of controls, minimally including variables used to stratify the nudges (course type and gender). Given the panel nature of our data, a subset of these can be replaced by individual fixed effects, and we include region-by-survey Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 month fixed effects to capture common changes in economic conditions. With respect to the relationship between platform usage and employment outcomes, we focus on three alternative estimators. One is the intent-to-treat (ITT) effect: yit = μ + δ j [Nudge j = 1]it + Xit γ + λt + εit , (2) j where δ j estimates the average effect of assignment to nudge j on outcome y. This captures both the effect associated with individuals induced to use the platforms on account of the nudge (compliers), as well as any effect associated with non-compliers, for whom the nudge was not salient. Second, since complier average treatment effects (CATE) are also often of stand-alone interest, espe- cially where non-compliance is material (Hernán and Hernández-Díaz 2012; Ye et al. 2014), we also apply a conventional instrumental variables estimator, denoted IV, in which assignment to a nudge is used as an instrument for platform usage. However, two drawbacks are associated with this approach—the exclusion restriction assumption is not directly testable and the IV estimator may have large bias and low precision in the presence of weak instruments (Young 2022). Both concerns are pertinent here. The exclusion restriction may be particularly fragile given the multiple pathways connecting the nudge to different dimensions of platform usage to final outcomes.11 Furthermore, as noted, we observe material two-way non-compliance (see table 1), implying assignment to treatment may be a poor indicator of eventual platform uptake. In light of this, a third approach invokes the assumption of principal ignorability (PI), which focuses on the different compliance strata to which each individual belongs. Following Frangakis and Rubin (2002) these strata are defined from the joint values of treatment assignment and uptake, namely always-takers (positive uptake regardless of assignment), never-takers (zero uptake regardless), compliers (uptake when assigned, no uptake when not assigned), and defiers (opposite of compliers). PI states that, conditional on included controls and treatment assignment, there is no unobserved factor that predicts both stratum membership and potential outcomes—i.e., once stratum membership is accounted for, potential outcomes can be compared within strata. In turn, any within-strata differences in outcomes between treatment and control groups should be attributable to treatment assignment.12 After accounting for controls, this effectively rules out the possibility that always-takers have higher gains from treatment than compliers, or that compliers have higher gains from treatment than never-takers. And while this may seem to run contrary to typical unobserved selection dynamics, we reiterate that our use of individual fixed effects should at least take into account all time-invariant unobserved effects that drive final outcomes. As elaborated in supplementary online appendix S4, we follow Ding and Lu (2017) and construct principal scores, focusing on the probability of being a complier, which we then apply as weights in a modified ITT regression analysis. This narrows attention to the principal stratum of interest—namely, compliers—placing higher (lower) weights on individuals most (least) likely to be compliers. In doing 11 The exclusion restriction also assumes no effect of treatment assignment on outcomes among non-compliers—i.e., for non-compliers we assume potential outcomes are the same regardless of assignment to treatment or control. This is equivalent to assuming the instrument only affects the outcome through treatment uptake (see De Chaisemartin 2017). 12 Stuart and Jo (2015) show that IV and PI estimates tend to “perform best under somewhat opposite scenarios” (also see Jo and Stuart 2009; Page et al. 2015; Feller, Mealli, and Miratrix 2017). Leveraging information on (predicted) compliance to sharpen treatment effect estimates has also been developed in other directions ( Follmann 2000; Joffe and Brensinger 2003), including weighted IV approaches (Huntington-Klein 2020). 152 Jones and Sen so, we leverage the fact that strata membership is partially observed. Those receiving an SMS nudge but with no observed platform usage are revealed never-takers, and individuals assigned to the control group but who nonetheless use a platform are always-takers. This approach represents a kind of weighted per-protocol analysis and relies on covariates, here including individual fixed effects, to control for any potential confounding between the sample selection criterion and outcomes (collider bias).13 Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 It can be noted that predicted compliance probabilities (weights) are only applied to individuals whose compliance status is ambiguous. For those whose compliance status is known, their observed (i.e., zero) compliance probability is used, which has the effect of focusing the PI analysis on the subsample of individuals who potentially responded to the encouragement treatment. As shown below, our own PI results are largely driven by this exclusion of the observed non-complier sample. 4. Results 4.1. Platform Uptake We start with the first-order issue of whether the nudges stimulated platform usage. Figure S2.6 illus- trates propensities to use the different platforms across rounds, classifying individuals by their eventual treatment assignment status. Following equation (1), table 2 reports results from separate OLS (linear probability) estimates of alternative metrics of platform usage on treatment assignment, controlling for experimental stratification covariates, selected baseline covariates, and region-by-month fixed effects. As per table 1, we note a consistent positive relationship running from nudges to usage. This relationship is strongest for the Biscate SMS, indicating a marginal effect of 0.47 for the external metric of usage (“Ext.”) on the same platform (column V). The equivalent marginal effect of the Emprego nudge is just 0.09 (col- umn I). Overall, exposure to the SMS nudges increased the average measure of usage on the targeted platforms by 0.08 and 0.37 for the Emprego and Biscate messages respectively (columns IV and VII)— i.e., relative to the control group, both nudges successfully stimulated platform usage. Yet, given a large share of nudge recipients did not go on to use the platforms, there was also significant non-compliance. Three further points emerge. First, since platform usage was not restricted to recipients of the nudges, we observe some uptake among the control group, representing a second form of non-compliance. For instance, among individuals assigned to the control group, uptake of Emprego reached an average of 0.17 by the final survey round and 0.05 on the Biscate platform. Second, there is indicative evidence of spillovers across the nudges.14 In particular, recipients of the Emprego SMS were not only more likely to use the Emprego platform, but also were marginally more likely to use the Biscate platform (by 0.08 points) than controls. As a consequence of possible spillovers across the two treatments, it is valuable to combine the measures of treatment assignment and uptake to capture their simultaneous effect. Column (VIII) of table 2 reports the analysis of uptake on this basis, where the outcome is now the maximum of the platform-specific usage means and the treatment is defined as having received any SMS nudge. This confirms a positive and large marginal effect of treatment assignment on platform usage—usage increased by 50 percent in the treatment group relative to controls. Last, there are systematic differences in platform usage according to baseline characteristics, partic- ularly gender. This is highlighted in panels (b) and (c) of the same table, which reruns the analysis for men and women separately. Within the control group, men are about 8 percentage points more likely to 13 This bias can emerge as observed (non)-compliance status is an outcome of both treatment assignment and uptake. Stratifying on or adding this term to the model opens a possible backdoor path to the outcome, which will be mate- rial to the extent that (non)-compliance status is associated with unobserved factors affecting outcomes ( Hernán and Hernández-Díaz 2012). 14 For reasons of inferential power associated with our sample size, we did not include a treatment arm combining both nudges. The World Bank Economic Review 153 Table 3. Estimates of Platform Effects on Employment Outcomes (I) (II) (III) (IV) (V) (VI) Combined treatments Separate treatments Estimator → ITT IV PI ITT IV PI Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 (a) Paid work: Any platform − 0.01 − 0.03 0.03 – – – (0.02) (0.12) (0.03) Emprego – – – − 0.01 − 0.13 − 0.00 (0.03) (0.37) (0.04) Biscate – – – − 0.00 − 0.01 0.05 (0.03) (0.08) (0.04) (b) Income: Any platform 0.02 0.08 0.11 – – – (0.05) (0.26) (0.07) Emprego – – – − 0.01 − 0.17 0.02 (0.06) (0.78) (0.09) Biscate – – – 0.04 0.10 0.21∗∗ (0.06) (0.17) (0.08) (c) Employment score: Any platform 0.03 0.12 0.11∗ – – – (0.05) (0.25) (0.06) Emprego – – – 0.01 0.06 0.09 (0.06) (0.79) (0.09) Biscate – – – 0.04 0.10 0.14∗ (0.06) (0.16) (0.08) Obs. 5,325 5,325 3,964 5,325 5,325 3,754 Source: Own estimates. Note: The columns and panels refer to separate regression estimates for versions of equation (2); the dependent variables are indicated by panels (a)–(c), of which (b) and (c) are regularized to mean 0 and standard deviation of 1 in the control group; (c) is the first principal component of all nine labor-market outcomes; income is IHS (inverse hyperbolic sine) transformed prior to regularization; ITT are intent-to-treat estimates; IV is an instrumental variables version of the complier-average treatment effect, in which the primary explanatory variable is platform usage (not the nudge); PI is a reweighted version of the ITT, based on compliance scores; only selected coefficients shown; all estimates contain both period (round × month × province) and individual fixed effects, as well as time-varying controls; standard errors (reported in parentheses) are clustered at the individual level. Significance: ∗ 10 percent, ∗∗ 5 percent, ∗∗∗ 1 percent. use either of the two platforms. Also, the marginal effect of the nudges is consistently stronger among men, by around 0.10 points, and especially with respect to the Emprego platform where the SMS nudge has almost no impact among women (columns I–IV). This suggests that existing gender differences in or even just perceptions of the labor-market moderate the impact of interventions to stimulate the uptake of digital platforms across men and women. 4.2. Average Labor-Market Effects Turning to final outcomes, table 3 reports results for each of our three main estimators with a focus on three headline employment metrics—having a paid job, monthly income (transformed using the inverse hyperbolic sine), and the overall employment score. In columns (I)–(III) we combine treatment assignment and uptake across the two platforms into single variables as before, while columns (IV)–(VI) keep the two platforms distinct. All estimates include individual fixed effects, which absorb the baseline covariates, as well as any unobserved time-invariant school-, course-, or individual-specific influences, period fixed effects, and a small set of time-varying covariates.15 The same estimates are shown visually in figs S2.7 to S2.9. 15 These are three subjective measures of the impact of COVID-19, pertaining to effects on their community, their family, and themselves. Inclusion of these covariates does not change our results. 154 Jones and Sen Two key points emerge. None of the ITT effects are distinguishable from zero, implying no statistically discernable differences between users assigned to the SMS nudges and the controls on average. This non- effect reflects the removal of both confounding associated with endogenous platform usage and material non-compliance with the nudges. At the same time, we recognize our estimates are not particularly sharp. For instance, the 95 percent confidence interval implied by the results in column (I) panel (a) range from Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 −0.05 to 0.04, which contains effect sizes that would be considered very reasonable for many active labor-market interventions (see below; also Card, Kluve, and Weber 2010). Second, the complier-average effects are somewhat ambiguous. The conventional IV estimates (columns II and V) also are not different from zero, but these suffer from even greater uncertainty, denoted by their large standard errors. For instance, the standard errors on the IV estimates for Emprego are more than 10 times the corresponding ITT estimates, and the range spanned by the 95 percent confidence interval in column (V) panel (a), namely [−0.37, 0.62], exceeds the range of the underlying outcome (unity).16 The PI estimates are comparatively more precise, but no more so than the ITT estimates. These estimates indicate that individuals who complied with the encouragement nudge achieved somewhat more positive employment outcomes, but such differences remain on the borderline of statistical significance. For example, there are no robust effects at the extensive margin of paid employment (panel a) and none of the PI effects associated with the Emprego nudge are different from zero at conventional significance levels. However, compliers with either (any) nudge recorded a marginal gain of 0.11 [iscate SMS treatment [−0.01, 0.29]. And compliers with the latter treatment also report higher incomes [0.05,0.37], an effect that is statisticallsignificant at the 5 percent level. Digging into the full set of outcomes underlying the overall employment score, table S2.2 reports coefficient estimates and tests of the null hypotheses, in each case repeating the main analyses from table 3 and adding results from a simple as-treated (denoted AT) estimator, in which observed (endogenous) platform usage is used as the main explanatory variable.17 Tests of the null hypothesis, that there are no effects from the platforms, are calculated both on a stand-alone basis and after adjustment for multiple hypothesis testing, where we apply the Benjamini and Hochberg (1995) procedure to control for the false discovery rate (calculated across all outcomes for each estimator). Figures S2.10 and S2.11 plot the corresponding point estimates and 95 percent confidence intervals associated with the any-platform and separate treatment assignments respectively. However, due to the large confidence intervals associated with the IV estimator, these results are not shown visually. Last, table S2.3 shows results for formal tests of whether the coefficient estimates for Emprego and Biscate are statistically different. Together, these findings largely reinforce earlier insights. While the as-treated effect estimates are generally in the positive domain, they are insignificant, and this applies to both the any-platform treatment assignments and the platform-specific results. Similarly, none of the ITT estimates are different from zero, the IV estimates are extremely uncertain, and the PI estimates lose significance after adjustment for multiple hypothesis testing. In light of the above, the power of our experiment merits further comment. Based on simulations in our pre-analysis plan, we estimated a minimum detectable effect of about 0.15 standard deviation units (or 7.5 percentage points for a binary outcome) for either of our treatments assuming 80 percent power at the 10 percent significance level with one observational round. These calculations apply to the ITT estimates and therefore would demand a much larger minimum complier-average detectable effect under significant non-compliance. Indeed, taking into account additional observations from multiple rounds, we approximate that complier-average effects would need to be in the region of 0.25 standard deviation 16 Hereafter, intervals stated in (square) brackets represent 95 percent confidence intervals. Note, individual fixed effects enter both the uptake and outcome equations under the IV estimator. 17 Prior to running these regressions we apply the inverse hyperbolic sine (IHS) transform to censored continuous variables (income, hours worked), which are (then) regularized to take a mean of 0 and standard deviation of 1 in the control group. Binary variables are not transformed. The World Bank Economic Review 155 units to be detected given the scale of non-compliance encountered in practice (ex post). Effectively, this means we are only powered to detect differential impacts on rates of employment in excess of around 10 percentage points, which would be a very large effect size especially given the light-touch nature of our SMS intervention. In this sense, our experiment provides comparatively clear insights into the (first-stage) impact of encouraging workers to use digital labor-market platforms, but we are in a weaker position to Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 draw confident conclusions about final outcomes. Further work on this topic should take into account the challenges of low take-up (compliance) in the design phase. 4.3. Robustness With respect to robustness, a first point to emphasize is that around 30 percent of all observations take a zero weight under the PI estimates. These refer to the observed never- or always-takers, who do not con- tribute to the complier-average effect by definition. All remaining observations are weighted by individual- specific estimated complier propensities, derived from separate non-parametric regressions for the treated and non-treated subsamples (for details see supplementary online appendix S4, where table S4.1 gives an overview of the estimated compliance scores). Specifically, in the PI estimates summarized in table 3, column (a) employs as weights the mean per-participant any-platform complier propensity multiplied by the original sample weights; in panel (b) the relevant platform-specific mean propensity is applied to individuals assigned-to-treatments, and we apply the any-platform measure to controls. Table S2.4 validates the sensitivity of these PI estimates to alternative choices for constructing complier weights. We begin simply by applying per-protocol weights, which amount to excluding all observed non-compliers from the sample. As shown in column (I), this reveals the predominant influence of this zero-weighting—i.e., per protocol weighting yields treatment-assignment coefficient estimates that closely approximate the PI results of table 3. Next we apply a shrinkage procedure to combine our estimated individual-specific compliance propensities (π ˆ c ) with the subgroup-specific (“true”) average propensities for 32 mutually exclusive groups ( p¯ c ), defined from pre-treatment characteristics. These are gender, course type, prior internship experience, and dichotomized versions of age and the outcome score in the first round. We calculate a shrunken propensity as γ π ˆ c,i + (1 − γ ) p ¯ c,i , and choose values for the shrinkage param- eter between 0.95 and 0.05. These results, reported in table S2.4 columns (II–VI), underline the previous point. As the shrinkage parameter falls, meaning revealed non-compliers take weights that approach sub- group means, the estimates of platform assignment essentially decline toward the ITT effect. Per se, this does not invalidate the PI procedure, but it does underscore the centrality of the principal ignorability as- sumption. And in this respect we highlight the use of individual fixed effects provides some comfort—by definition, these partial out any pre-treatment factors that confound principal stratum membership and potential outcomes. A further aspect of robustness concerns spillovers in treatment assignment and platform usage among peers, such as via information sharing. To address this, we calculate leave-one-out peer group averages for these two variables, defining these groups as individuals of the same gender attending the same baseline survey data collection session (typically, students of the same or closely related study course), stratified by follow-up round—similar to course×round fixed effects. We then include these variables as additional covariates in our regressions. Effectively, this means we control for the share of peers assigned to each nudge or using each platform in each round. Reported in table S2.5, this extension does not alter our core results. However, we see the IV estimates for any treatment assignment (panel a) and the Biscate nudge (panel b) are now highly similar in magnitude to those of the PI estimates, but remain insignificant at conventional levels. 156 Jones and Sen 4.4. Gender Heterogeneity Following earlier discussion we pursue a gender heterogeneity analysis. This is further motivated by stark differences in employment outcomes between men and women in our sample. For instance, as shown in table S2.1, within the control group we see men are at least 10 percentage points more likely to be working than women, and such differences are even wider among students of manual courses—here, 42 percent Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 of men in the control group obtained paid work, but just 17 percent of women. Such gender gaps are consistent with other studies of the Mozambican labor market, which contend that young women face material trade-offs between work and household reproductive activities, meaning that economic returns to education predominantly accrue to men (Gradín and Tarp 2019; Bischler et al. 2022). Reflecting what Tvedten (2011) describes as a “deeply patriarchal society,” women face much greater difficulties in being able to participate in formal (full-time) work. To investigate whether gender differences moderate the impacts of digital labor-market platforms, we rerun our previous estimates interacting the treatment nudges with the individual’s gender. Table 4 reports the results, replicating the structure of the previous analysis, and fig. 2 illustrates the ITT and PI estimates for men and women, showing the net effects associated with each of the two platforms. Albeit with fairly low precision, the majority of these estimates indicate the platforms were moderately less beneficial for women. For instance, based on our preferred PI estimates, for the average male complier the marginal impact of platform usage on the overall employment score is 0.18 [0.04,0.32] standard deviation units (see panel c), but the same effect for women is close to zero at −0.01 [−0.17, 0.16]. Column (VI) indi- cates this difference is driven principally by the Emprego platform, where both the ITT and PI estimates indicate a statistically significant lower effect of Emprego for women compared to men. However, for Biscate, we note a positive and significant effect on undertaking paid employment among men, but not women. To triangulate these findings we conclude with three complementary exercises. First, we review gender differences across the set of outcomes underlying the employment score. As per tables S2.7 and S2.8 (also figures S2.12 and S2.13), gender differences in treatment effects are evident across many outcomes but generally fail to meet conventional thresholds for statistical significance, particularly after adjustment for multiple hypothesis testing. Nonetheless, using the PI estimates, we note that job quality is higher in a number of dimensions among male compliers with the Emprego nudge, but not for women, substantiating the higher overall employment score obtained by men compared to women on this platform.18 Second, we extend our analysis to selected secondary outcomes, namely those pertaining to job-search behavior. As summarized in table 5 (also figures S2.14 and S2.15), these suggest that one channel that may help explain gender gaps in the effects of the platforms on employment outcomes is their differential impact on job search.19 In particular, while the reservation wage of men does not seem to alter with plat- form usage, that of women increases significantly (by around 0.10 standard deviations), especially among those exposed to the Biscate nudge. At the same time, male users of either platform appear somewhat more likely to be actively seeking work, but spend less time on job search when nudged to use Emprego— suggesting some efficiency gains in job search when using the platform. In contrast, none of these effects obtain among women. Third, to externally validate the presence of significant gender differences in employment outcomes on digital platforms, we take advantage of anonymized individual-level data from Biscate. As in Jones and Manhique (2024), where the same data are described in detail, the platform records information on basic individual characteristics (registration data) as well as work-related outcomes—namely, the number of times they were contacted by a client and agreed a task. Since data at the individual level are both sparse 18 Further gender differences emerge if we also consider interactions with course type, but our inferential power here becomes even weaker, meaning any conclusions cannot be drawn confidently. 19 For reference, table S2.6 summarizes the average effects for the same secondary outcomes. The World Bank Economic Review 157 Table 4. Estimates of Platform Effects on Employment Outcomes by Gender (I) (II) (III) (IV) (V) (VI) Combined treatments Separate treatments Estimator → ITT IV PI ITT IV PI Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 (a) Paid work: Any 0.01 0.02 0.06 – – – (0.03) (0.11) (0.04) Any × female − 0.04 − 0.14 − 0.08∗ – – – (0.03) (0.12) (0.04) Emprego – – – 0.00 − 0.07 0.02 (0.04) (0.31) (0.05) Emprego × female – – – − 0.02 − 0.25 − 0.06 (0.05) (0.53) (0.07) Biscate – – – 0.02 0.03 0.10∗∗ (0.04) (0.09) (0.04) Biscate × female – – – − 0.05 − 0.10 − 0.11∗ (0.04) (0.15) (0.06) (b) Income: Any 0.04 0.14 0.14∗ – – – (0.06) (0.24) (0.08) Any × female − 0.06 − 0.15 − 0.06 – – – (0.07) (0.24) (0.09) Emprego – – – 0.00 − 0.12 0.04 (0.08) (0.66) (0.11) Emprego × female – – – − 0.02 − 0.18 − 0.07 (0.09) (1.12) (0.15) Biscate – – – 0.07 0.16 0.22∗∗ (0.08) (0.21) (0.10) Biscate × female – – – − 0.08 − 0.17 − 0.03 (0.09) (0.33) (0.12) (c) Employment score: Any 0.06 0.21 0.18∗∗ – – – (0.06) (0.23) (0.07) Any × female − 0.08 − 0.24 − 0.19∗∗ – – – (0.07) (0.26) (0.09) Emprego – – – 0.10 0.42 0.23∗∗ (0.08) (0.66) (0.10) Emprego × female – – – − 0.19∗∗ − 1.81 − 0.39∗∗ (0.10) (1.21) (0.15) Biscate – – – 0.03 − 0.05 0.16∗ (0.07) (0.19) (0.09) Biscate × female – – – 0.02 0.44 − 0.05 (0.09) (0.35) (0.13) Obs. 5,325 5,325 3,964 5,325 5,325 3,754 Source: Own estimates. Note: The columns and panels refer to separate regression estimates for versions of equation (2); the dependent variables are indicated by panels (a)–(c), of which (b) and (c) are regularized to mean 0 and standard deviation of 1 in the control group; the employment score is the first principal component of all nine labor-market outcomes; income is IHS (inverse hyperbolic sine) transformed prior to regularization; ITT are intent-to-treat estimates; IV is an instrumental variables version of the complier-average treatment effect, in which the primary explanatory variable is platform usage (not the nudge); PI is a reweighted version of the ITT, based on compliance scores; only selected coefficients shown; all estimates contain both survey period (round × month × province) and individual fixed effects, as well as time-varying controls; standard errors (reported in parentheses) are clustered at the individual level. Significance: ∗ 10 percent, ∗∗ 5 percent, ∗∗∗ 1 percent. 158 Jones and Sen Figure 2. Heterogeneous Effect Estimates for Headline Employment Outcomes, by Gender. Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 Source: Own estimates. Note: The figures plot point estimates and 95 percent confidence intervals for heterogeneous treatment effects based on interaction terms, as per table 4; estimators are ITT and PI; continuous dependent variables are regularized to mean 0 and standard deviation of 1 in the control group. The World Bank Economic Review 159 Table 5. Estimates of Platform Effects on Job-Search Behavior by Gender (I) (II) (III) (IV) (V) (VI) Combined treatments Separate treatments Estimator → ITT IV PI ITT IV PI Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 (a) Seeking work: Any 0.04 0.13 0.10∗∗ – – – (0.03) (0.11) (0.03) Any × female − 0.06∗∗ − 0.19 − 0.09∗∗ – – – (0.03) (0.12) (0.04) Emprego – – – 0.02 0.11 0.10∗∗ (0.04) (0.29) (0.05) Emprego × female – – – − 0.02 0.16 − 0.07 (0.05) (0.62) (0.07) Biscate – – – 0.05 0.13 0.13∗∗ (0.03) (0.09) (0.04) Biscate × female – – – − 0.10∗∗ − 0.27∗ − 0.12∗∗ (0.04) (0.15) (0.05) (b) Search hours: Any − 0.19∗ − 0.71∗ − 0.20 – – – (0.11) (0.41) (0.13) Any × female 0.14 0.21 0.03 – – – (0.11) (0.41) (0.15) Emprego – – – − 0.34∗∗ − 2.07∗ − 0.49∗∗ (0.14) (1.15) (0.18) Emprego × female – – – 0.43∗∗ 2.72 0.49∗ (0.16) (1.88) (0.28) Biscate – – – − 0.08 0.07 0.02 (0.13) (0.34) (0.16) Biscate × female – – – − 0.12 − 1.14∗∗ − 0.20 (0.15) (0.56) (0.20) (c) Reservation wage: Any 0.01 0.08 0.00 – – – (0.03) (0.11) (0.04) Any × female 0.07∗∗ 0.32∗∗ 0.07 – – – (0.03) (0.12) (0.04) Emprego – – – 0.01 0.09 − 0.03 (0.04) (0.30) (0.05) Emprego × female – – – 0.04 0.28 0.02 (0.04) (0.58) (0.08) Biscate – – – 0.01 0.05 − 0.01 (0.04) (0.10) (0.05) Biscate × female – – – 0.10∗∗ 0.28∗ 0.12∗ (0.04) (0.16) (0.06) Obs. 5,325 5,325 3,964 5,325 5,325 3,754 Source: Own estimates. Note: The columns and panels refer to separate regression estimates for versions of equation (2); the dependent variables are indicated by panels (a)–(c), of which (b) and (c) are both IHS (inverse hyperbolic sine) transformed and regularized to mean 0 and standard deviation of 1 in the control group; IV is an instrumental variables version of the complier-average treatment effect, in which the primary explanatory variable is platform usage (not the nudge); PI is a reweighted version of the intent-to-treat (ITT) estimate, based on compliance scores; only selected coefficients shown; all estimates contain both survey period (round × month × province) and individual fixed effects, as well as time-varying controls; standard errors (reported in parentheses) are clustered at the individual level. Significance: ∗ 10 percent, ∗∗ 5 percent, ∗∗∗ 1 percent. 160 Jones and Sen and vary with the time of registration, we aggregate observations by province and profession on a weekly basis and investigate systematic determinants of these outcomes, which we convert into rates (e.g., number of contacts per 100 registered workers). Table S2.10 reports our results, which explicitly include gender- and profession-type interaction terms, capturing how outcomes vary with the composition of workers or area of work (see columns b and c).20 Controlling for a rich set of fixed effects, we find that a marginal Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 increase in the share of women registered as available to work in (male-dominated) manual professions is associated with significantly higher contact and agreement rates. But the opposite tendency applies to women in service-related jobs, where the majority of registered female workers are in fact registered.21 In other words, women generally face greater barriers to finding work than men on Biscate, but there is some evidence they can achieve comparatively better outcomes in specific niches where they stand out on account of their gender. 5. Conclusion This paper engaged with a growing literature on how search frictions impede labor-market outcomes for youth in developing countries. Adding to previous studies that have mostly considered transport and screening costs in formal labor markets, we compared the impacts on jobs outcomes of digital matching platforms in distinct market segments—for informal tasks and formal jobs. Focusing on recent gradu- ates from TVET colleges in Mozambique, we embedded a randomized encouragement design within a tracer survey and invited participants to register on one of two platforms: Biscate, a portal to find infor- mal freelancers for tasks, or Emprego, a conventional website posting formal employment opportunities. Our encouragement nudge induced only moderate platform uptake in the treatment groups relative to controls, which occurred alongside non-zero uptake in the control group. Thus, we complemented intent- to-treat estimates with econometric methods to estimate complier-average treatment effects based on an assumption of principal ignorability. In keeping with a handful of prior studies from Asia, we did not detect a material robust causal effect of the platforms on labor-market outcomes on average. Neither rates of employment, job quality, nor wage incomes altered materially as a result of platform usage. These results apply to both platforms and obtain across different estimators, but there is weak evidence of a small positive effect of being nudged to use the informal tasks platform (Biscate). In part, these results reflect limitations in our statistical power due to material non-compliance, meaning that smaller but nonetheless meaningful effects cannot be reliably detected. Looking beyond averages, a heterogeneity analysis nonetheless revealed important gender differences. We found men were more responsive to the encouragement nudge and that platform usage moderately benefited the average male (complier) but not the average female, especially with respect to an overall metric of employment quality. Such gender differences are corroborated by evidence that women using the platforms increased their reservation wage, unlike men who were more likely to be actively seeking work. Two complementary interpretations can be made of these results. The first concerns market satura- tion or excess labor supply. A number of studies highlight the importance of credible signalling on digital platforms (also Carranza et al. 2022)—in saturated market segments, individuals may find it hard to obtain work when they cannot easily differentiate themselves from other candidates. Unless digital plat- forms can provide credible information on worker skills or quality, they may do little to reduce matching frictions. In the context of Mozambique, where there are significant demographic pressures and a weak 20 Individuals register on Biscate within predefined professional categories as well as within a specific location (province). Search functionality follows this classification and we distinguish between manual and service-type professions. 21 On average, 32 percent of registered workers in service-related professions on Biscate are women, falling to 6 percent in manual professions. The World Bank Economic Review 161 formal labor market, as well as the coincidence of our survey with the COVID-19 pandemic (Jones and Manhique 2024), this may well explain our average null effects and weak-to-moderate results for men (only). A broader implication is that digital jobs platforms are unlikely to be a general panacea for youth un(der)employment in low-income Africa, at least while the supply of good quality employment openings is limited. Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 Second, in line with Afridi et al. (2022), our findings suggest that existing gender norms associated with work, such as those that constrain women from pursuing formal jobs, are plausibly reproduced by digital platforms. The point is that new technologies often operate in domains where potential users face heterogeneous pre-existing opportunity sets, which in turn influence their capacity to access and benefit from such technologies. This consideration is pertinent to Mozambique, where formal jobs are highly male dominated and women carry the burden of household reproduction. Thus, in expanding the role of digital platforms, care must be taken to understand existing labor-market constraints and gender-based dynamics. Further research on how and when digital platforms can reduce rather than reproduce gender gaps in employment is warranted. Data Availability Statement The data underlying this article will be shared on reasonable request to the corresponding author. References Abebe, G., S. Caria, M. Fafchamps, P. Falco, S. Franklin, and S. Quinn. 2021. “Anonymity or Distance? Job Search and Labour Market Exclusion in a Growing African City.” Review of Economic Studies 88(3): 1279–310. Abebe, G., S. Caria, M. Fafchamps, P. Falco, S. Franklin, S. Quinn, and F. Shilpi. 2021. “Matching Frictions and Distorted Beliefs: Evidence from a Job Fair Experiment.” G2LMLIC Working Paper 49, IZA Institute of Labor Economics. Adams-Prassl, A., K. Hara, K. Milland, and C. Callison-Burch. 2023. “The Gender Wage Gap in an Online Labor Market: The Cost of Interruptions.” Review of Economics and Statistics forthcoming: 1–23. DOI: 10.1162/rest_a_01282. Afridi, F., A. Dhillon, S. Roy, and N. Sangwan. 2022. “Social Networks, Gender Norms and Women’s Labor Supply: Experimental Evidence Using a Job Search Platform.” STEG Working Paper WP044, Centre for Economic Policy Research (CEPR), Structural Transformation and Economic Growth (STEG). A Anwar, M.. 2022. “Platforms of Inequality: Gender Dynamics of Digital Labour in Africa.” Gender & Development 30(3): 747–64. Bandiera, O., A. Elsayed, A. Heil, and A. Smurra. 2022. “Economic Development and the Organisation of Labour: Evidence from the Jobs of the World Project.” Journal of the European Economic Association 20(6): 2226–70. Beam, E. A. 2016. “Do Job Fairs Matter? Experimental Evidence on the Impact of Job-Fair Attendance.” Journal of Development Economics 120: 32–40. Benjamini, Y., and Y. Hochberg. 1995. “Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing.” Journal of the Royal Statistical Society: Series B (Methodological) 57(1): 289–300. Bhuller, M., G. B. Dahl, K. V. Løken, and M. Mogstad. 2020. “Incarceration, Recidivism, and Employment.” Journal of Political Economy 128(4): 1269–324. Bischler, J., E.-M. Egger, P. Jasper, and I. Manhique. 2022. “Determinants of Gender Gaps in Youth Employment in Urban Mozambique.” WIDER Working Paper 164/2022, UNU-WIDER, Helsinki, Finland. Card, D., J. Kluve, and A. Weber. 2010. “Active Labour Market Policy Evaluations: A Meta-Analysis.” Economic Journal 120(548): F452–77. Carranza, E., R. Garlick, K. Orkin, and N. Rankin. 2022. “Job Search and Hiring with Limited Information about Workseekers’ Skills.” American Economic Review 112(11): 3547–83. Carranza, E., and D. Mckenzie. 2024. “Job Training and Job Search Assistance Policies in Developing Countries.” Journal of Economic Perspectives, 38(1): 221–244. 162 Jones and Sen Chade, H., J. Eeckhout, and L. Smith. 2017. “Sorting through Search and Matching Models in Economics.” Journal of Economic Literature 55(2): 493–544. Chakravorty, B., A. Y. Bhatiya, C. Imbert, M. Lohnert, P. Panda, and R. Rathelot. 2021. “Impact of COVID-19 Crisis on Rural Youth: Evidence from a Panel Survey and an Experiment.” GLO Discussion Paper Series 909, Global Labor Organization (GLO). Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 Chung, H., and M. Van der Horst. 2018. “Women’s Employment Patterns after Childbirth and the Perceived Access to and Use of Flexitime and Teleworking.” Human Relations 71(1): 47–72. De Chaisemartin, C. 2017. “Tolerating Defiance? Local Average Treatment Effects without Monotonicity.” Quanti- tative Economics 8(2): 367–96. Ding, P., and J. Lu. 2017. “Principal Stratification Analysis Using Principal Scores.” Journal of the Royal Statistical Society. Series B (Statistical Methodology) 79(3): 757–77. Feller, A., F. Mealli, and L. Miratrix. 2017. “Principal Score Methods: Assumptions, Extensions, and Practical Con- siderations.” Journal of Educational and Behavioral Statistics 42(6): 726–58. A Follmann, D.. 2000. “On the Effect of Treatment among Would-Be Treatment Compliers: An Analysis of the Multiple Risk Factor Intervention Trial.” Journal of the American Statistical Association 95(452): 1101–09. Frangakis, C. E., and D. B. Rubin. 2002. “Principal Stratification in Causal Inference.” Biometrics 58(1): 21–29. Franklin, S. 2018. “Location, Search Costs and Youth Unemployment: Experimental Evidence from Transport Subsi- dies.” Economic Journal 128(614): 2353–79. Gradín, C., and F. Tarp. 2019. “Gender Inequality in Employment in Mozambique.” South African Journal of Eco- nomics 87(2): 180–99. Hamal, P., and R. Huijsmans. 2022. “Making Markets Gendered: Kathmandu’s Ride-Sharing Platforms through a Gender Lens.” Gender, Place & Culture 29(5): 670–92. Hernán, M. A., and S. Hernández-Díaz. 2012. “Beyond the Intention-to-Treat in Comparative Effectiveness Research.” Clinical Trials 9(1): 48–55. Huntington-Klein, N. 2020. “Instruments with Heterogeneous Effects: Bias, Monotonicity, and Localness.” Journal of Causal Inference 8(1): 182–208. ILO. 2021. World Employment and Social Outlook 2021: The Role of Digital Labour Platforms in Transforming the World of Work. Geneva: International Labour Organization. ———. 2022a. Global Employment Trends for Youth 2022: Investing in Transforming Futures for Young People. Geneva: International Labour Office. ———. 2022b. World Employment and Social Outlook – Trends 2022. Geneva: International Labour Organization. Jayachandran, S. 2021. “Social Norms as a Barrier to Women’s Employment in Developing Countries.” IMF Economic Review 69(3): 576–95. Jo, B., and E. A. Stuart. 2009. “On the Use of Propensity Scores in Principal Causal Effect Estimation.” Statistics in Medicine 28(23): 2857–75. Joffe, M. M., and C. Brensinger. 2003. “Weighting in Instrumental Variables and g-Estimation.” Statistics in Medicine 22(8): 1285–303. Jones, S., and I. Manhique. 2024. “Digital Labour Platforms as Shock Absorbers: Evidence from the COVID-19 Pandemic in Mozambique.” Journal of African Economies forthcoming: ejae002. Jones, S., and R. Santos. 2020. “Can Digital Labour Market Platforms Get Africa’s Youth Working? Evidence from Mozambique.” Pre-analysis plan, AEA RCT Registry. https://www.socialscienceregistry.org/trials/6465. ———. 2022. “Can Information Correct Optimistic Wage Expectations? Evidence from Mozambican Job-Seekers.” Journal of Development Economics 159: 102987. Jones, S., R. Santos, and G. Xirinda. 2021. “Survey on the School-To-Work Transition of Technical and Vocational Training Graduates in Mozambique.” Final report, UNU-WIDER. https://www.wider.unu.edu/publication/survey - school- work- transition- technical- and- vocational- training- graduates- mozambique. Kelley, E. M., C. Ksoll, and J. Magruder. 2024. “How Do Digital Platforms Affect Employment and Job Search? Evidence from India.” Journal of Development Economics: 166: 103176. Kroft, K., and D. G. Pope. 2014. “Does Online Search Crowd Out Traditional Search and Improve Matching Effi- ciency? Evidence from Craigslist.” Journal of Labor Economics 32(2): 259–303. The World Bank Economic Review 163 Marbach, M., and D. Hangartner. 2020. “Profiling Compliers and Noncompliers for Instrumental-Variable Analysis.” Political Analysis 28(3): 435–44. Mulalic, I., J. N. Van Ommeren, and N. Pilegaard. 2014. “Wages and Commuting: Quasi-Natural Experiments’ Evi- dence from Firms That Relocate.” The Economic Journal 124(579): 1086–1105. Page, L. C., A. Feller, T. Grindal, L. Miratrix, and M.-A. Somers. 2015. “Principal Stratification: A Tool for Un- Downloaded from https://academic.oup.com/wber/article/39/1/143/7667623 by WORLDBANK THIRDPARTY user on 05 February 2025 derstanding Variation in Program Effects across Endogenous Subgroups.” American Journal of Evaluation 36(4): 514–31. Pallais, A. 2014. “Inefficient Hiring in Entry-Level Labor Markets.” American Economic Review 104(11): 3565–99. Rani, U., R. Castel-Branco, S. Satija, and M. Nayar. 2022. “Women, Work, and the Digital Economy.” Gender & Development 30(3): 421–35. Rodríguez-Modroño, P., A. Pesole, and P. López-Igual. 2022. “Assessing Gender Inequality in Digital Labour Platforms in Europe.” Internet Policy Review 11(1): 1–23. Rud, J. P., and I. Trapeznikova. 2021. “Job Creation and Wages in Least Developed Countries: Evidence from Sub- Saharan Africa.” Economic Journal 131(635): 1331–64. Stanton, C. T., and C. Thomas. 2016. “Landing the First Job: The Value of Intermediaries in Online Hiring.” Review of Economic Studies 83(2): 810–54. Stuart, E. A., and B. Jo. 2015. “Assessing the Sensitivity of Methods for Estimating Principal Causal Effects.” Statistical Methods in Medical Research 24(6): 657–74. Tvedten, I. 2011. “Gender Equality and Development in Mozambique: Background Article for the World Development Report 2012.” Unpublished report, World Bank. online Van Den, Berg, and G. J. 1999. “Empirical Inference with Equilibrium Search Models of the Labour Market.” Eco- nomic Journal 109(456): 283–306. Wheeler, L., R. Garlick, E. Johnson, P. Shaw, and M. Gargano. 2022. “LinkedIn (to) Job Opportunities: Experimental Evidence from Job Readiness Training.” American Economic Journal: Applied Economics 14(2): 101–25. Ye, C., J. Beyene, G. Browne, and L. Thabane. 2014. “Estimating Treatment Effects in Randomised Controlled Trials with Non-compliance: A Simulation Study.” BMJ Open 4(6): e005362. Young, A. 2022. “Consistency without Inference: Instrumental Variables in Practical Application.” European Eco- nomic Review 147: 104112.