Directed search, wages, and nonwage amenities: Evidence from an online job board

Escudero, Verónica; Liepmann, Hannah; Vergara, Damián Working Paper Directed search, wages, and nonwage amenities: Evidence from an online job board ILO Working Paper, No. 136 Provided in Cooperation with: International Labour Organization (ILO), Geneva Suggested Citation: Escudero, Verónica; Liepmann, Hannah; Vergara, Damián (2025) : Directed search, wages, and nonwage amenities: Evidence from an online job board, ILO Working Paper, No. 136, ISBN 978-92-2-041393-7, International Labour Organization (ILO), Geneva, https://doi.org/10.54394/YWML9238 This Version is available at: https://hdl.handle.net/10419/315005 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ XDirected Search, Wages, and NonWage Amenities: Evidence from an Online Job Board Authors / Verónica Escudero, Hannah Liepmann, Damián Vergara March / 2025 ILO Working Paper 136 © International Labour Organization 2025 Attribution 4.0 International (CC BY 4.0) This work is licensed under the Creative Commons Attribution 4.0 International. See: https:// creativecommons.org/licenses/by/4.0/. The user is allowed to reuse, share (copy and redistribute), adapt (remix, transform and build upon the original work) as detailed in the licence. The user must clearly credit the ILO as the source of the material and indicate if changes were made to the original content. Use of the emblem, name and logo of the ILO is not permitted in connection with translations, adaptations or other derivative works. Attribution – The user must indicate if changes were made and must cite the work as follows: Escudero, V., Liepmann, H., Vergara, D. Directed Search, Wages, and Non-Wage Amenities: : Evidence from an Online Job Board. ILO Working Paper 136. Geneva: International Labour Office, 2025.© ILO. Translations – In case of a translation of this work, the following disclaimer must be added along with the attribution: This is a translation of a copyrighted work of the International Labour Organization (ILO). This translation has not been prepared, reviewed or endorsed by the ILO and should not be considered an official ILO translation. The ILO disclaims all responsibility for its content and accuracy. Responsibility rests solely with the author(s) of the translation. Adaptations – In case of an adaptation of this work, the following disclaimer must be added along with the attribution: This is an adaptation of a copyrighted work of the International Labour Organization (ILO). This adaptation has not been prepared, reviewed or endorsed by the ILO and should not be considered an official ILO adaptation. The ILO disclaims all responsibility for its content and accuracy. Responsibility rests solely with the author(s) of the adaptation. Third-party materials – This Creative Commons licence does not apply to non-ILO copyright materials included in this publication. If the material is attributed to a third party, the user of such material is solely responsible for clearing the rights with the rights holder and for any claims of infringement. Any dispute arising under this licence that cannot be settled amicably shall be referred to arbitration in accordance with the Arbitration Rules of the United Nations Commission on International Trade Law (UNCITRAL). The parties shall be bound by any arbitration award rendered as a result of such arbitration as the final adjudication of such a dispute. For details on rights and licensing, contact: [email protected]. For details on ILO publications and digital products, visit: www.ilo.org/publns. ISBN 9789220413920 (print), ISBN 9789220413937 (web PDF), ISBN 9789220413944 (epub), ISBN 9789220413951 (html). ISSN 2708-3438 (print), ISSN 2708-3446 (digital) https://doi.org/10.54394/YWML9238 The designations employed in ILO publications, which are in conformity with United Nations practice, and the presentation of material therein do not imply the expression of any opinion whatsoever on the part of the ILO concerning the legal status of any country, area or territory or of its authorities, or concerning the delimitation of its frontiers or boundaries. See: www.ilo. org/disclaimer. The opinions and views expressed in this publication are those of the author(s) and do not necessarily reflect the opinions, views or policies of the ILO. Reference to names of firms and commercial products and processes does not imply their endorsement by the ILO, and any failure to mention a particular firm, commercial product or process is not a sign of disapproval. Information on ILO publications and digital products can be found at: www.ilo.org/researchand-publications ILO Working Papers summarize the results of ILO research in progress, and seek to stimulate discussion of a range of issues related to the world of work. Comments on this ILO Working Paper are welcome and can be sent to escuder[email protected]. Authorization for publication: Richard Samans, Director, Research Department ILO Working Papers can be found at: www.ilo.org/global/publications/working-papers Suggested citation: Escudero, V., Liepmann, H., Vergara, D. 2025. Directed Search, Wages, and Non-Wage Amenities: : Evidence from an Online Job Board, ILO Working Paper 136 (Geneva, ILO). https://doi. org/10.54394/YWML9238 01 ILO Working Paper 136 Abstract We leverage rich data from a prominent online job board in Uruguay to assess directed search patterns in job applications, focusing on posted wages and advertised non-wage amenities. We find robust evidence of directed search based on posted wages in the cross-section, with stark heterogeneity by occupation: the wage-application correlation is driven by vacancies attached to lower-skill occupations, with applications to vacancies attached to higher-skill occupations show - ing no responsiveness to posted wages. By applying text analysis to the job ads, we elicit advertised non-wage amenities and find evidence of directed search based on non-wage amenities. Applications to vacancies attached to lower-skill occupations are consistent with lexicographic application preferences: amenities predict applications to these vacancies only when wages are not posted. Finally, we exploit industry-by-occupation minimum wage variation to demonstrate that the observed occupational heterogeneity in directed search patterns is supported by quasi-experimental difference-in-differences estimates of the impact of wages on job applications. JEL codes: E24, J31, J32, J62, J63 Keywords: Directed Search, Vacancies, Wages, Non-Wage Amenities, Minimum Wages About the authors Verónica Escudero joined the Research Department of the ILO in 2008 and today she is Chief of the Skills, ALMPs and Policy Evaluation Team. Between March 2021 and February 2023, she served as a Visiting Scholar with CEGA (Center for Effective Global Action) at the University of California Berkeley. She is a PhD specialized on labour and development economics and applied microeconomics. Her current research focuses on assessing the effectiveness of labour market and social policies on job quality and social conditions. More recently, she has been exploring topics related to the skills necessary to foster effective transitions to decent work with a focus on lowand middle-income countries, through the use of online data on vacancies and applications to labour portals. She holds a PhD in Economics from Paris School of Economics and the École des Hautes Études en Sciences Sociales (EHESS). Hannah Liepmann joined the Research Department of the International Labour Organization in 2018, where she works as an Economist in the Skills, Active Labour Market Policies, and Policy Evaluation Team. As an empirical labour economist, she is particularly interested in studying how labour market and social protection policies as well as phenomena of structural change affect the integration of marginalized groups into quality employment. Hannah obtained her PhD in Economics from Humboldt-University Berlin and she is an IZA Research Fellow. Damián Vergara is an Assistant Professor in the Department of Economics at the University of Michigan with research interests in public and labor economics. In the 2023-2024 academic year, he was a Postdoctoral Research Associate at the Industrial Relations Section at Princeton University. Vergara obtained a B.A. and an M.A. in Economics from the Universidad de Chile and a Ph.D. in Economics from UC Berkeley. 02 ILO Working Paper 136 Abstract 01 About the authors 01 XIntroduction 04 Related literature 07 Structure of the paper 08 X1 General Context, Data, and Descriptive Statistics 09 1.1 General context 09 1.2 Data 09 1.3 Descriptive statistics 11 Vacancies 11 Applicants 12 Applications 12 Amenities 12 X2 Cross-Sectional Facts on Job Applications 14 2.1 How diverse are application portfolios? 14 Number of applications 14 Diversity in applications 14 2.2 Cross-sectional patterns of directed search 17 Applications and wages 17 Occupational heterogeneity 18 Applicant-level heterogeneity 20 2.3 The role of non-wage amenities 21 Posted wages and amenities 21 Applications and amenities 23 Applicant-level heterogeneity 24 2.4 Summary of findings and discussion 26 X3 The Causal Effect of Wages on Applications 27 3.1 Setting and data 27 Collective Bargaining Agreements 27 CBAs data 28 Economic significance of the minimum wage across occupations 29 3.2 Empirical strategy 29 Table of contents 03 ILO Working Paper 136 Estimating equations 30 3.3 Application effects of the minimum wage 31 Robustness checks and within-vacancy design 32 Heterogeneity by applicant characteristics 33 3.4 Additional results 33 Vacancies and openings 33 Advertised non-wage amenities 33 Vacancy requirements 34 XConclusions 35 References 36 Figures and Tables 42 A. Methodology for Creating Variables from Free Text Entries 62 A.1 Skills 62 A.2 Occupations 64 A.3 Amenities 65 B. Additional Figures and Tables 70 Appendix Bibliography 93 Acknowledgements 96 04 ILO Working Paper 136 XIntroduction How responsive are job seekers to the characteristics of vacancies? Unpacking the “black box” of job applications informs about the presence of labor market frictions and helps to assess key assumptions in related theoretical work, for example, about random versus directed search, wage posting versus wage bargaining, or the role of non-wage amenities. Understanding the job application process is particularly relevant given recent documentation of imperfect information and beliefs in the labor market, from both workers and employers (Cullen, 2024; Jäger et al., 2024). Moreover, as argued by Holzer et al. (1991), job queuing behavior suggests the existence of ex-ante rents in the labor market. Hence, job seekers’ responses to changes in the characteristics of posted vacancies can shed light on the degree to which the documented industryand firm-level wage premia constitute evidence of rents in the labor market. Despite its importance, the empirical study of job applications is challenging because most datasets record equilibrium outcomes which, by definition, are only observed once the job application process is completed. To overcome this challenge, researchers have switched gears to gather direct information on the application process. Hall and Krueger (2012) and Krueger and Mueller (2016) pioneered using survey data on workers and job seekers. More recently, economists have started using vacancy-level data from private online job boards to better understand how firms advertise jobs and recruit workers and how job seekers search and make application decisions (e.g., Banfi and Villena-Roldan, 2019; Marinescu and Wolthoff, 2020; Skoda, 2022; Arnold et al., 2023; Batra et al., 2023). This paper builds on this latter literature and uses data from a large online job board in Uruguay to study directed search patterns in job applications (that is, the extent to which job seekers direct their search toward vacancies with specific attributes), focusing on the role of posted wages and advertised non-wage amenities. The data comes from BuscoJobs (BJ), a prominent online job search platform that operates in more than 30 countries. In Uruguay, BJ covers a broad set of industries and occupations and is estimated to contain around 60% of total online private sector vacancies in the country (Escudero et al., Forthcoming). We have access to data on vacancies, applicants, and applications for the period 2011-2020, which we link using unique identifiers of applicant profiles and vacancies. On top of the complete application portfolio, applicant profiles contain information on gender, age, employment status, employment histories, education, and training. Vacancies contain information on the number of positions they seek to fill, formal requirements, and firm and industry identifiers. Also, 20% of vacancies post a monthly wage. Important for our analysis, we have access to the full job ad text, which is processed using Natural Language Processing (NLP) techniques to elicit the following additional variables: the skills required by vacancies and the occupations vacancies seek to recruit (Escudero et al., Forthcoming); and the non-wage amenities advertised in the job post (Adamczyk et al., Forthcoming). The analysis proceeds in two parts. The first part develops a cross-sectional analysis that confirms and extends the main findings of Banfi and Villena-Roldan (2019) and Marinescu and Wolthoff (2020). The analysis shows that vacancies attached to lower-skill occupations receive more applications when they post higher wages or advertise non-wage amenities, however, there is a lexicographic application pattern as the effect of amenities on applications vanishes in the subset of vacancies that post a wage. On the contrary, applications to vacancies attached to higher-skill occupations do not react to posted wages but increase when non-wage amenities are advertised. The second part of the analysis leverages the fact that Uruguay implements Collective Bargaining Agreements (CBAs) that dictate and frequently adjust minimum wages, which vary at the industry-by-occupation level. We exploit this feature to complement the cross-sectional analysis with 05 ILO Working Paper 136 causal differences-in-differences estimates of wage effects on job applications. We find that minimum wage hikes increase applications to vacancies attached to lower-skill occupations, with no effect on vacancies attached to higher-skill occupations. Hence, the occupational heterogeneity documented in the cross-section is corroborated in the quasi-experimental exercise. To preview our analysis in more detail, the cross-sectional analysis is structured in three exercises. First, we characterize application portfolios at the applicant level and explore whether they are diversified or concentrated in a few industries or occupations. We find substantial heterogeneity in the number of applications per job search spell across applicants. We also document that application portfolios are diversified. Job seekers who submit multiple applications in a given quarter rarely concentrate their applications within a specific industry and/or occupation. Instead, workers tend to apply for vacancies that span a wide range of industries and occupations. For example, when applicants submit 5 applications in a given quarter, their applications span, on average, 4.2 2-digit industries, 3.5 1-digit industries, and 2.8 1-digit occupations. This qualitative pattern remains consistent regardless of the number of applications made. This set of findings suggests that workers do not exhibit strong attachments to an occupation and, especially, an industry at the time of application, implying that they possibly consider other job attributes when choosing the vacancies they apply for, making directed search patterns feasible. The second cross-sectional exercise explicitly explores directed search based on posted wages. We first replicate Banfi and Villena-Roldan (2019) and Marinescu and Wolthoff (2020) main finding of a positive and significant correlation between posted wages and vacancy-level applications once appropriate skill controls (in our case, occupations) are included. The main contribution of this section, however, is to document a stark heterogeneity by occupation in the wage-application elasticity. We find that for vacancies attached to a subset of occupations, which we label as lower-skill occupations (clerical support, services and sales, plant and machine operators, and elementary occupations), the elasticity of applications to posted wages is large, significant, and highly robust to the inclusion of controls and sample selections. On the contrary, for the vacancies attached to the remaining occupations, which we label as higher-skill occupations (managers, professionals, technicians and associate professionals, and craft workers), the relationship between applications and posted wages is completely absent. This finding is consistent with wage posting being more prevalent in lower-skill occupations and wage bargaining and individual offer tailoring being more prevalent in higher-skill occupations (e.g., Hall and Krueger, 2012; Caldwell and Harmon, 2019; Lachowska et al., 2022; Caldwell et al., 2024) since posted wages may provide different information for applicants depending on their occupation, thus mediating the application responsiveness. The documented heterogeneity is also consistent with recent evidence on employers being more likely to use aggregate information to set wages (that is, engage in “salary benchmarking”) when positions are attached to low-skill occupations (Cullen et al., 2024). In this exercise, we also take advantage of the applicant-level data and explore whether the responsiveness of applications to posted wages varies with applicant characteristics. Both a vacancy-level analysis and an application-level analysis show that applications made by male, employed, older, college-educated, and skilled job seekers are significantly more responsive to wages than applications made by female, unemployed, young, non-college-educated, and unskilled job seekers, respectively. We also find that applicants with presumably worse labor market prospects (female, unemployed, young, non-educated, and unskilled) display negative wage elasticities when applying to vacancies attached to higher-skill occupations. This finding is consistent with models of directed search where workers trade-off wages with job search spell length (e.g., Moen, 1997) and models with on-the-job search where worse outside options may encourage workers to apply to low-wage jobs to climb the job ladder in future job transitions (e.g., Burdett and Mortensen, 1998; Postel-Vinay and Robin, 2002a,b). 12 ILO Working Paper 136 Panel (a) of Table 1 also shows the shares of vacancies that specify requirements for applicants. 14% of vacancies require a vocational training certificate, while 21% require a college degree, and 19% of vacancies require knowledge of a language other than Spanish (in most cases, English). These requirements are directly specified by firms in related entries when posting a vacancy. Meanwhile, other requirements are specified in the open text of job ads, which are elicited using NLP techniques as described in Section 1.2. We observe that 80% of vacancies require at least one cognitive skill. Likewise, 83% of vacancies require at least one socio-emotional skill, and 38% require at least one manual skill. Applicants Panel (b) of Table 1 presents descriptive statistics of the applicants registered in the BJ platform. We identify 698,880 profiles in the 2010-2020 period, of which 410,955 are denoted as “active” profiles, i.e., individuals who made at least one application in the period October 2011 to September 2020.8 To get a sense of the order of magnitude, the total population in Uruguay was estimated to be 3,530,912 in 2020, of which 2,067,384 (59%) was between 20 and 64 years old (INE, 2021). This implies that the total number of profiles created between 2010 and 2020 represents approximately 34% of the working-age population in Uruguay in 2020.9 Among the active profiles, the mean number of applications made between October 2011 and September 2020 is 39.7 and the corresponding median is 11. Among active applicants, 55% are female. 11% report having a vocational training degree, and 16% have completed a college degree. The median applicant was born in 1990. While the overall educational structure resembles that of the national labor force, BJ applicants are slightly more likely to be college graduates.10 They also include a disproportionate number of younger workers (Escudero et al., Forthcoming). Applications We identify 16,320,466 applications for vacancies made between October 2011 and September 2020. Panel (c) of Table 1 shows that 42% of applications are made by individuals who report being employed at the time of the application. The average age at the time of application is relatively young at 27.7 years with moderate dispersion. Finally, based on the open-text descriptions of current and previous jobs, we estimate that 31% report having performed cognitive tasks, 45% report having performed socio-emotional tasks, and 13% report having performed manual tasks. Amenities Table 2 shows descriptive statistics for the amenities advertised in our sample of vacancies. 45% of the vacancies in our sample advertise at least one of the 5 amenities described in Section 2.2. The average number of amenities advertised is 0.7, which can be decomposed as 55% of vacancies advertising zero amenities, 27% advertising one, 12% advertising two, 4% advertising three, and less than 2% advertising four or five amenities. The share of vacancies advertising amenities is larger among vacancies that do not post wages (46% versus 39%). Table 2 also shows that 8244,960 profiles report no application in the 2010-2020 period, while 42,965 only made applications between January 2010 and September 2011 and/or October 2020 and December 2020. 9As an alternative benchmark, 114,392 individuals made at least one job application on the platform in 2020, which corresponds to 6% of the population aged 20 to 64. 10 Of the national labor force, 14% completed a vocational degree and 11% a college degree. Thus, the share without such qualifications is large for both BJ applicants and the overall labor force (see Escudero et al., Forthcoming, where cited figures are 2010-2020 averages based on the Uruguayan household survey). 13 ILO Working Paper 136 the three most commonly advertised amenities are “human capital development”, “working in teams”, and “work environment/impact on society”, which are featured in 23%, 19%, and 16% of vacancies, respectively. “Bonuses and commissions” and “schedule flexibility” are advertised in 7% and 5% of the vacancies, respectively. 14 ILO Working Paper 136 X2 Cross-Sectional Facts on Job Applications Having described the setting and the data, we proceed with the cross-sectional analysis. We perform three exercises. First, we explore how diverse application portfolios are in terms of industries and occupations. Second, we explore whether posted wages affect applications. Third, we explore the role of advertised non-wage amenities in the application process. We pay particular attention to heterogeneity analyses at the applicant, vacancy, and amenity levels during the analysis. 2.1 How diverse are application portfolios? To study application portfolios, we analyze the universe of applications made to the vacancies considered in our analysis. To proxy groups of applications made in the same job search spell (applicants may search for jobs at multiple stages in their careers, thus applying to jobs in different job search spells), we consider an applicant ID-by-quarter-by-year as a unit of observation and focus on “active spells”, i.e., applicant ID-by-quarter-by-year combinations where job seekers make at least one application. If actual application spells are longer than a quarter, our measure will underestimate the number of applications by search spell. This strategy leads to 1,668,348 applicant-by-spell observations with at least one application. On average, an active profile makes applications in 4.1 different quarters between 2011 and 2020. There is, however, substantial heterogeneity. 34.4% of the applicant profiles are active only in one quarter, 37.4% between 2 and 4 different quarters, 16.0% between 5 and 8 quarters, 6.4% between 9 and 12 quarters, and 5.8% in 13 or more different quarters.11 Number of applications Figure 2 shows the distribution of the number of applications at the applicant-by-spell level, again considering active applicant-by-spell combinations with a positive number of applications. Panel (a) shows wide variation in the number of applications across applicants. While 23% of applicants with positive applications make a unique application, 51% of applicants submit between 2 and 10 applications in a given spell, and 13% submit between 11 and 20 applications. Only 3% of applicants submit more than 50 applications in a given spell (not shown in the histogram). Panels (b), (c), and (d) show the distributions separately by employment status, gender, and educational attainment. Distributions look remarkably similar across demographic groups, especially with respect to gender. While employed and college-educated applicants tend to make fewer applications on average than unemployed and non-college-educated applicants, they still show wide dispersion in the number of applications. Diversity in applications Having established that applicants are heterogeneous in the number of applications made by active spell, we then explore whether applications made by a given applicant in a given quarter 11 Applicants who are active only for one quarter may be different than the average applicant. For example, they may enter the BJ website but then use it less actively as they quickly find employment. In Figure B.2 of Appendix B, we show that the distribution of the number of applications indeed changes slightly when excluding applicants who are active only during one quarter. Compared to Figure 2, the average number of applications during active spells tends to increase. However, in Figure B.4 of Appendix B, we also show that the results of this subsection are robust to excluding the applicant IDs that apply for jobs only during one quarter. 15 ILO Working Paper 136 tend to target vacancies in specific industries or occupations or if, instead, their applications are diversified across industries and occupations. This analysis can inform about the extent of directed search in the labor market: if workers are strongly attached to particular industries and occupations and, therefore, their behavior is less responsive to wage differentials across industries and occupations, we would expect to see their job applications concentrated within narrow categories of vacancies. We explore diversification in application portfolios using the following statistic. Let i index observations (active applicant ID-spell combinations) with Ni the total applications made by the applicant in the spell. Each application goes to a vacancy attached to a group g ∈ G, with #G the number of different possible groups. For example, G may be the set of 2-digit industry codes, g a particular 2-digit industry, and #G the number of different 2-digit industries. Let #gi ∈ {1, ..., min{#G, Ni}} be the number of groups spanned by the Ni applications of applicant i.12 For example, if Ni = 10, #gi = 5 means that the 10 applications span 5 different 2-digit industries. When Ni = 1, #gi is mechanically 1. When Ni > 1, the upper bound of gi is given by min{#G, Ni}. We measure diversification with the quantity: When D(N) = 1, applications are not diversified: all are made to the same group of vacancies. When D(N) = N, applications are completely diversified: all are made to vacancies that belong to different groups. This implies that the distance between (N, D(N)) and the 45-degree line can be used to visually diagnose the extent of diversification in application portfolios, taking into account the heterogeneity in the number of applications documented above. We study diversification focusing on four different groups of vacancies. We first consider a narrow definition of vacancy groups that share their 2-digit industry code and their 1-digit occupation code, thus employing a stricter definition of a possibly relevant labor market.13 If industry-by-occupation cells constitute an accurate definition of the relevant local labor market of the applicant, we should expect job seekers to make the majority of their applications to vacancies in the same industry-by-occupation cell.14 We also consider broader group definitions of relevant labor markets: 2-digit industry codes alone, 1-digit industry codes alone, and 1-digit occupation codes alone. Figure 3 shows the results. We focus on applicant-spell observations making 10 or fewer applications (Ni ∈ {1, ..., 10}). The black dotted curve is the 45-degree line. Figure B.3 of Appendix B shows results for Ni ∈ {1, ..., 50}, which displays a similar pattern.15 Two aspects of the figure are worth highlighting. First, when considering the narrower group definition (blue curve, 2-digit industry by 1-digit occupation cell), the levels of diversification are substantial. For example, D(2) = 1.96, which means that almost everyone who applies to 2 vacancies applies to vacancies in 2 different industry-by-occupation cells. While D(N)/N decreases with the number of applications, it remains large across the distribution of N. Individuals making 5 and 10 applications span 4.6 and 8.7 industry-by-occupation cells, respectively. This result implies that job seekers rarely target 12 Formally, #gi can be thought of as the cardinality of the partition of Ni in the space of G. 13 We follow the categorizations of ISCO 08 for occupations and ISIC Rev. 4 for industries. 14 The usual definition of a local labor market also considers a geographical dimension (e.g., Manning and Petrongolo, 2017). We disregard this dimension since more than 50% of the country lives in the metropolitan area of Montevideo, the capital city, and the data does not allow us to do a more granular analysis within the city. 15 When applicants make too many applications, D(N) is more likely to be mechanically affected by #G, affecting the interpretability of D(N) in the tail of the distribution. 16 ILO Working Paper 136 industry-by-occupation cells when making applications. Second, while mechanically smaller, diversification remains large when considering broader groups (2-digit industries, 1-digit industries, and 1-digit occupations alone). Applicants who make 2 applications span, on average, 1.90 2-digit industries, 1.79 1-digit industries, and 1.66 1-digit occupations. Applicants who make 5 applications span, on average, 4.2 2-digit industries, 3.5 1-digit industries, and 2.8 1-digit occupations. Applicants who make 10 applications span, on average, 7.1 2-digit industries, 5.1 1-digit industries, and 3.8 1-digit occupations.16 The fact that industries seem to be more diversified than occupations is worth highlighting. It implies that it is more accurate to think that, when applying, workers fix occupations and arbitrage industries than the other way around, suggesting that directed search is plausible. For the very least, the analysis rejects the hypothesis that applicants target labor markets defined by narrow industry-by-occupation cells, which would limit their sensitivity to wage differentials across industries. In that spirit, the analysis suggests that industry-wage differentials (Krueger and Summers, 1988; Card et al., 2024) cannot be rationalized by workers having strong attachment to particular industries. On the contrary, it suggests that industry wage differentials may give form to a job ladder. A possible caveat of these results is that, conditional on making several applications, the vacancy offer distribution at a given point in time may be limited, thus preventing job seekers from implementing “non-diversified” application portfolios. For example, at a given point in time, there may be few vacancies (maybe one or none) associated with a particular industry-by-occupation cell. In that sense, the spike at 1 in Figure 2 may partially reflect an aversion to diversification for a subset of applicants. However, the fact that we observe a non-trivial share of job seekers making multiple applications and that, conditional on making multiple applications, job seekers apply to a wide range of vacancies in terms of industries and occupations, shows that significant numbers of applicants are, in fact, diversifying their applications. Diversification seems indeed substantial even conditional on only making 2 applications. Then, if this concern is driving the results, we should observe all applicants making a unique application, a pattern that is strongly rejected in Figure 2.17 As a final test for “willingness to diversify,” we leverage the fact that, for the majority of employed job seekers, we observe the 1-digit occupation of their current job and the occupation attached to the vacancies they apply to. Then, we can observe the share of on-the-job applications that are made to vacancies attached to the same occupation as the current job. Figure 4 shows the results split by occupation of the current job and number of applications made in the job search spell. As a benchmark, if applications were made randomly, the share of applications targeted to vacancies attached to the same occupation would match the distribution of posted vacancies displayed in Panel (b) of Figure 1. Figure 4 shows that job seekers in all occupations display shares larger than the benchmark suggested by Figure 1, which implies that workers apply more often to vacancies attached to their current occupations. However, the figure reveals that job seekers who apply for jobs while employed are also willing to apply to vacancies attached to other occupations, suggesting that diversification is plausible. This pattern is observed even for applicants 16 This result is not exclusively explained by differences in #G, since D(N) remains far below the upper bounds. The number of 2-digit industry by 1-digit occupation cells observed in the vacancy data is 504. The number is 70, 14, and 8 for 2-digit industries, 1-digit industries, and 1-digit occupations, respectively. 17 Figure B.5 of Appendix B explores for heterogeneities by applicant characteristics (employment status, gender, education, and job search spell length). The figures suggest that all subgroups of applicants are diversifying their applications. Per the concern described above, differences in diversification by applicant characteristics may be partially reflected by the differential distributions in the number of applications observed in Figure 2. 17 ILO Working Paper 136 who make only 1 application in a given job search spell. This result is consistent with recent findings in Altmann et al. (2024) and Fluchtmann et al. (Forthcoming). 2.2 Cross-sectional patterns of directed search The fact that job seekers have diversified application portfolios suggests they may direct their search based on characteristics other than occupation and industry. This subsection explores directed search patterns based on posted wages. As mentioned previously, among our sample of 77,874 vacancies, 15,835 (20.3%) include a salary range in the posted ad. On average, vacancies that post a wage receive 21% more applications (with a median of 29%) relative to those that do not include wage information. While the distribution of industries and occupations is similar between vacancies that post and do not post wages (see Figure 1), the difference in applications suggests that the decision to post a wage may be endogenous. In Section 3, we therefore complement the cross-sectional analysis with quasi-experimental results to provide further ground for a causal interpretation of the effects of wages on applications. Applications and wages We start by non-parametrically exploring the relationship between log applications and log posted wages, pooling all vacancies that post a wage in our dataset. Figure 5 shows different binscatter plots that vary in the controls considered. Panel (a) shows the raw relationship between log applications and log posted wages. The plot shows a noisy and inverse U-shaped relationship: vacancies that post very low or very high wages tend to receive fewer applications. Panel (b) shows that the same relationship is observed when excluding the 3% of outlier vacancies that received more than 1,000 applications and controlling for 2-digit industry fixed effects, year fixed effects, and the advertised non-wage amenities in the vacancy. As stressed by Banfi and Villena-Roldan (2019) and Marinescu and Wolthoff (2020), the cross-sectional relationship may be spurious when not properly controlling for the skills associated with the job tasks. Panels (c) and (d) add 1-digit and 2-digit occupation fixed effects, respectively, and suggest that, with the exception of the vacancies at the very top of the posted wage distribution, the relationship between applications and posted wages becomes positive, suggesting the presence of a within-occupation directed search pattern in a wide range of the posted wage distribution. To summarize these patterns in terms of cross-sectional wage-application elasticities, we run OLS regressions of the following type: where Appj is the number of applications per opening for vacancy j, wj is the posted wage of vacancy j, and Xj are vacancy-level controls. We cluster standard errors at the 2-digit industry level. Panel (a) of Table 3 shows the estimate of α under different sets of controls, resembling the analysis in Figure 5. Column (1) shows the raw correlation, which is positive but small and not statistically significant. Column (2) excludes outliers, includes industry and year fixed effects, and controls for the advertised amenities. Including this set of controls has a small effect on the coefficient but slightly increases precision. Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively, generating an increase in the estimated coefficient that is statistically significant at conventional levels. The cross-sectional application-wage elasticity in these columns is 0.17 and 0.19, respectively. Column (5) excludes the vacancies at the top 5% of the 18 ILO Working Paper 136 posted wage distribution. Consistent with Figure 5, omitting the upper tail increases the elasticity to 0.33. Finally, as a robustness check, Column (6) leverages the fact that several firms in the platform post multiple vacancies and, therefore, considers only vacancies posted by firms with 10 or more posted vacancies and includes firm fixed effects. While it is not clear whether firm fixed effects are good controls (directed search may reflect job ladders between firms), it is reassuring that the elasticity remains positive and significant, with a value of 0.21.18 We note that our analysis replicates the main conclusions in Banfi and Villena-Roldan (2019) and Marinescu and Wolthoff (2020): directed search arises after including appropriate controls for the skill attached to the vacancy, which we approximate with occupation codes. Our estimated elasticities are smaller than the ones estimated in the aforementioned papers, possibly because they use job titles as the skill control which are substantially narrower than 1-digit or 2-digit occupational codes. The sensibility of the estimated elasticity to the included controls, however, provides a similar narrative in qualitative terms.19,20 In most of the exercises that follow, we report results for the same sets of controls and sample refinements. Given the lessons from related literature and the results of Panel (a) in Table 3, however, we designate the specification of Column (3) (no outliers, 2-digit industry fixed effects, year fixed effects, advertised amenities, and 1-digit occupation fixed effects) as our preferred specification. The choice of 1-digit over 2-digit occupation codes rests solely on the fact that 2-digit codes are not available for all vacancies and, therefore, using 1-digit codes increases the sample size. Occupational heterogeneity One caveat of the analysis above is that it pools all vacancies when estimating the cross-sectional wage-application relationship. It could be the case, however, that different occupations react differently to posted wages. For example, findings in Hall and Krueger (2012), Caldwell and Harmon (2019), Lachowska et al. (2022), and Caldwell et al. (2024) suggest that wage bargaining is more prevalent in higher-skilled occupations, a feature that could mediate how job seekers attached to different occupations interpret and react to posted wages in online job ads. To explore for occupational heterogeneities, we replicate the analysis separately by 1-digit occupation categories. Figure 6 presents binscatter plots for the relationship between applications and posted wages by occupation. These figures exclude outliers and include industry fixed effects, year fixed effects, and controls for advertised amenities. The data reveals the existence of two groups of occupations that display opposite patterns. Panel (a) shows results for vacancies attached to clerical support, services and sales, plant and machine operators, and elementary occupations. We denote this group of occupations as lower-skill and refer to vacancies attached to these occupations as “lower-skill vacancies”. Vacancies in this group exhibit a monotone and positive relationship between applications and posted wages. Panel (b) shows results for vacancies attached to managers, professionals, technicians and associate professionals, and craft workers. We denote 18 The 77,874 vacancies are posted by 6,214 firms. 2,341 firms only post one vacancy, 2,578 firms post between 2 and 9 vacancies, and 1,295 firms post 10 or more vacancies. 2,682 firms are responsible for the 20% of vacancies that post wage. 19 Table B.3 of Appendix B replicates Panel (a) of Table 3 using different definitions of posted wage. Panel (a) uses the midpoint of the salary range. Panel (b) uses the midpoint of the salary range but excludes vacancies with ranges larger than 50% of the midpoint. Panel (c) uses the maximum of the salary range. In these cases, and consistent with Banfi and Villena-Roldan (2019) and Marinescu and Wolthoff (2020), elasticities are negative in the absence of skill controls but become positive when adding the occupation fixed effects. When using the midpoint of the salary range, the qualitative pattern of Table 3 is confirmed and the resulting elasticities are significant. When using the maximum of the salary range, the same qualitative pattern is observed but with smaller and non-significant estimates. These results suggest that the minimum of the salary range is presumably more relevant to job seekers to decide on applications relative to the maximum. 20 1.6% (260) of vacancies advertising a wage had zero applications and are, therefore, excluded from the main analysis. Table B.2 of Appendix B shows that a Poisson model that includes the vacancies with zero applications yields similar results. 19 ILO Working Paper 136 this group of occupations as higher-skill and refer to vacancies attached to these occupations as “higher-skill vacancies”. The relationship between applications and posted wages is essentially flat for this group of occupations. Panel (c) reproduces the analysis after grouping lowerand higher-skill occupations into the two broad groups. The lower-skill group exhibits a clear positive correlation between applications and posted wages, whereas the higher-skill group displays no such relationship.21 Table B.4 in Appendix B presents estimates of equation (2) separate by 1-digit occupation group, confirming the patterns displayed in Figure 6. Panel (b) of Table 3 summarizes the results by replicating Panel (a) of Table 3 using a model with interactions: where LS and HS account for lowerand higher-skill occupation, respectively. When Xj does not include 1-digit or 2-digit occupation fixed effects (Columns (1) and (2)), the regression controls for 1{Occj ∈ LS}. The results are remarkably stable across columns and confirm the pattern documented in Figure 6. Vacancies attached to lower-skill occupations consistently display a positive and significant elasticity of applications to posted wages, with larger magnitudes closer to the values reported in Banfi and Villena-Roldan (2019) and Marinescu and Wolthoff (2020). On the contrary, vacancies attached to higher-skill occupations show no significant relationship between posted wages and applications. In our preferred specification (Column (3)), the point estimates are  α LS= 0.41 and  α HS = -0.07, compared to the estimated  α = 0.17 from the regression with no interactions. One possible explanation for this pattern is that higher-skill vacancies impose more requirements on applicants in terms of formal qualifications or skills, which could prevent job seekers from applying to high-wage higher-skill vacancies. In fact, Table B.5 of Appendix B shows that requirements are more prevalent in higher-skill vacancies. Table B.6 of Appendix B, however, shows that the absence of directed search in higher-skill vacancies holds whether we restrict the sample to vacancies that post or do not post requirements. This is true in terms of formal requirements (see Panels (a) and (c), which focus on vocational training, college degrees, and/or language requirements) and skill requirements (see Panels (b) and (d), which pertain to cognitive, socio-emotional, and/or manual skills). Conversely, we continue to find evidence of directed search among vacancies attached to lower-skill occupations independent of vacancy requirements. Yet, for vacancies attached to these occupations, the responsiveness to posted wages is stronger when no formal requirements are posted (Panels (a) and (c)).22 21 We refer to the two data-driven occupational groups as lowerand higher-skilled for the sake of exposition clarity and consistency with existing economic literature (e.g., Kunst et al., 2022; Montobbio et al., 2023) while acknowledging the limitations of skill-based categories based on broad occupation codes. At a high level, this categorization is aligned with the ISCO-08 guidelines (ILO, 2012), which classifies the four 1-digit occupations in our lower-skilled group at the lowest skill levels 1 and 2, and three of the four 1-digit occupations in our higher-skilled group (managers, professionals, and technicians and associate professionals) at the highest skill levels 3 and 4. In the ISCO-08 guidelines, skill levels are determined based on the complexity and range of tasks and duties typically associated with an occupation, as well as the level of formal education required to perform those tasks. This classification system relies on a broad generalization of tasks and duties typically performed within an occupation without accounting for variation in task complexity across different jobs within the same occupation or between countries. Moreover, it places greater emphasis on formal educational qualifications despite the importance of other types of learning, for example, on-the-job (Konings and Vanormelingen, 2015; Attanasio et al., 2011; Alfonsi et al., 2020). It is worth noting that our higher-skill group includes craft and related trades workers which ISCO-08 classifies at skill level 2. One possible explanation for why craft and trades workers exhibit similar application patterns to the rest of occupations in the higher-skill group is the level of wages. This occupation typically commands a higher average salary compared to other occupations classified at a similar ISCO skill level 2 (see Table B.14 of Appendix B). 22 For the sub-sample of vacancies not requiring any cognitive, socio-emotional or manual skills, some of the coefficients are more imprecisely estimated, which may be due to the smaller sample size. 20 ILO Working Paper 136 Applicant-level heterogeneity Finally, we leverage our applicant-level data and test whether directed search patterns are heterogeneous by applicant characteristics. We proceed in two ways. First, we estimate equation (3) using applications from particular groups of applicants as dependent variables. Table 4 presents the results. Regressions exclude outliers and include 2-digit industry fixed effects, year fixed effects, amenity controls, and 1-digit occupation fixed effects. While the larger responsiveness in vacancies attached to lower-skill occupations relative to higher-skill occupations is seen across all groups of applicants, point estimates reveal substantial heterogeneity by group of applicants. Columns (1) and (2) of Panel (a) show results for female and male applicants, respectively. Male applicants are substantially more responsive to posted wages than female applicants. While the lower-skill elasticity is 0.63 for male applicants, female applicants reveal a non-significant lower-skill elasticity of 0.14 and a negative higher-skill elasticity of -0.27. Columns (3) and (4) of Panel (a) provide a similar comparison between employed and unemployed applicants, with employed applicants showing a much larger responsiveness to posted wages in lower-skilled vacancies (  αLS = 0.61) than unemployed applicants (  α LS = 0.28). The latter group also reveals a negative higher-skill elasticity of -0.22. Columns (5) and (6) of Panel (a) compare applications from job seekers aged 25 or less with applications from job seekers older than 25. Younger applicants show a precisely estimated zero lower-skill elasticity and a large negative higher-skill elasticity of -0.50. On the contrary, the estimated elasticities for older applicants resemble the findings for male and employed applicants (  αLS = 0.74 and  αHS = 0.20). Panel (b) shows heterogeneities by applicants’ education and skill level. Columns (1) to (3) of Panel (b) reveals larger responsiveness for applicants with tertiary education, especially for job seekers with a college degree in lower-skilled occupations (  αLS = 0.83). Columns (4) to (6) show that applicants with cognitive, socio-emotional, and manual skills, are also more responsive to posted wages, both in lowerand higher-skill vacancies.23 It is noteworthy that the groups of applicants with presumably worse labor market prospects exhibit negative higher-skill elasticities. The finding for these groups is consistent both with models of directed search where workers trade-off wages with job search spell length (e.g., Moen, 1997) and with models with on-the-job search where the lack of outside employment options may encourage workers to apply to low-wage jobs with the aim of climbing the job ladder in future job transitions (e.g., Burdett and Mortensen, 1998; Postel-Vinay and Robin, 2002a,b). The particularly negative response observed for the younger applicants may also reflect that these workers internalize in their application choices the importance of labor market experience to access higher-paying jobs. As a second exploration for applicant-level heterogeneity in directed search behavior, we use application-level data to test whether applicant demographics predict the log posted wage of the vacancy they are applying to conditional on being a vacancy that posts wage. Let i index applicants and j index vacancies. Then, for all applications made to vacancies that post a wage, we estimate: where Ageij is applicant i’s age when applying to vacancy j; and Femalei , Employedij , Voc.Trn.i , College i , Cogn.Sk. ij , Soc.Sk. ij , and Man.Sk. ij are indicator variables taking the value of 1 if applicant 23 Not all vacancies receive applications that span the complete distribution of applicants observables. As a consequence, the number of vacancies considered in each regression is not constant. As displayed in Table B.7 of Appendix B, similar results are obtained from a Poisson regression model that includes vacancies with zero applications. 21 ILO Working Paper 136 i is, respectively, female, employed when applying to the vacancy, has a vocational training degree, has a college degree, and reports having cognitive, socio-emotional, and manual skills when applying.24 As above, Xj contains vacancy-level controls which may include industry fixed effects, year fixed effects, advertised amenities, and occupation fixed effects. Standard errors are clustered at the applicant level. One advantage of this approach relative to the vacancy-level exercise reported in Table 4 is that, by controlling simultaneously for all applicant characteristics, it can better isolate the partial correlation of a specific attribute. If applicant characteristics are correlated with each other, the exercise above may be picking similar variation across columns, spuriously attributing results to particular characteristics. Table B.8 of Appendix B shows the results. The analysis is consistent with the results of the vacancy-level analysis of Table 4. Estimates are all significant and remarkably stable across specifications, suggesting that applicant-level heterogeneity is not driven by differential sorting to vacancies. The preferred specification (Column (3)) suggests the following point estimates, with little variation across columns. Relative to male job seekers, female job seekers apply to vacancies that post 5.7% lower wages. Relative to the unemployed, employed applicants apply to vacancies that post 5.2% higher wages. Being one year older predicts a 0.6% higher posted wage. Relative to applicants with no tertiary education, applicants with vocational training and a college degree apply to vacancies that post, on average, 3.9% and 10.8% higher wages, respectively. Likewise, applicants with cognitive skills, socio-emotional skills, and manual skills, apply to vacancies with 5.3%, 0.7%, and 1.5% higher posted wages, respectively. 2.3 The role of non-wage amenities The analysis above focused on the relationship between applications and wages, in some cases controlling for the amenities advertised in the vacancy. It does not, however, explore the concrete role amenities play for job seekers, which may yield a deeper understanding of the previously shown heterogeneities in the results on wages. Then, in this third exercise, we explore correlations that inform the role of non-wage amenities in the application process. We start by exploring correlations between posted wages and amenities. We then analyze how amenities correlate with applications and show heterogeneities by amenity, vacancy characteristics, and applicants’ characteristics. Posted wages and amenities To document correlations between advertised amenities and posted wages, we estimate OLS regressions of the form: where A includes bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development, and Am j a = 1{Vacancy j advertises amenity a}. We also consider regressions that only include Amj = 1{Vacancy j advertises at least 1 of the 5 amenities}. The potential set of controls is the same as in the previous subsection, and 24 In our data, formal education indicators are time-invariant. However, skills variables are time-variant since they are built from current employment and, in the case of unemployed individuals, employment histories. 28 ILO Working Paper 136 CBAs are negotiated at the group level. Groups correspond to broad economic industries. Each group bargains over one or more CBAs, depending on the number of subgroups considered. The objective of having different contracts within a group is to accommodate economic differences between sub-industries, although all CBAs within a group are jointly bargained and, in some cases, exhibit little within-group heterogeneity. Each group has discretion to define the occupations (i.e., categories) that will be subject to specific minimum wages in the bargained CBA. Figure 7 provides additional information about the CBAs. Panel (a) displays the number of groups and CBAs (subgroups) by year. In 2005, CBAs were negotiated in 20 groups. This increased to 24 groups in 2008 when CBAs became available for economic activities that were historically excluded from the wage councils, such as the domestic workers group and three groups representing activities of the rural economy. Within each group, there are the above-mentioned subgroups that negotiate different CBAs. There were 172 subgroups negotiating CBAs in 2005, which reached 221 with the incorporation of the previously excluded groups in 2008, covering virtually all private sector employees. Since then, the increase in the number of subgroups is explained by a reorganization within groups rather than an increase in coverage. Panel (b) shows the distribution and evolution of the number of wage floors defined within CBAs. As explained before, groups have autonomy to define the categories that will be affected by the sectoral minimum wages defined in the CBAs. The number of categories corresponds to the number of occupations with a fixed minimum wage in the CBA. There is substantial dispersion in the number of categories considered. In a typical year, a CBA in the 25th percentile defined 6 different minimum wages, while a CBA in the 75th percentile defined between 25 and 30 different minimum wages. Some groups defined more than a hundred categories, which explains the distance between the median and the average. The number of categories within CBAs is very stable over time. Finally, Panel (c) shows that, among the 96,598 minimum wage changes that we identify in the raw CBA data, more than 90% happened either in January or July. This feature will be important for our empirical strategy below. CBAs data We rely on information of the industry-by-occupation minimum wages defined in each tripartite negotiation and recorded in the CBAs. After each bargaining round, each group defines nominal wages and biannual adjustments that are valid until the next bargaining round. CBAs and the corresponding list of industry-by-occupation minimum wages are public information. Therefore, we use digitized minimum wage data collected from each round’s contract. One caveat when merging the CBAs data with the BJ data is that groups and categories that define industry-by-occupation minimum wages do not map one-to-one to the standardized codes available in the BJ data. Therefore, we manually assign 2-digit ISIC codes to each subgroup, and 1-digit ISCO occupation codes to each category within the contract. One problem with this approach is that the imputed codes can be broader or narrower than the standardized codes. For example, within the group “Commerce”, there is a CBA for the subgroup “Stores”, for which several 2-digit ISIC codes apply. Likewise, within the group “Food and beverages manufacturing”, there are different CBAs for “Wheat Mills” and “Rice Mills”, which can be associated to the same 2-digit industry code. A similar issue occurs with the categories within each CBA. One minimum wage can be associated with several 1-digit occupation codes, and several minimum wages can be associated with the same 1-digit occupation code. Since it is not possible to precisely attach specific minimum wages to vacancies because of this multiplicity problem, we build measures of exposure to minimum wage changes at the vacancy level by computing summary statistics of all minimum wages that can be associated to a specific 29 ILO Working Paper 136 industry-occupation combination. Then, we code whether, on a given date, there is a change in the computed statistic.28 Under this strategy, non-exposed vacancies are vacancies whose industry-occupation combination either cannot be associated with a minimum wage in any CBA, or vacancies for which we can associate a minimum wage but it is not changing in the specific period. The resulting cells are defined at the 2-digit industry level and the 1-digit occupational level. Economic significance of the minimum wage across occupations Our causal analysis aims to assess whether we can replicate the occupational heterogeneity in directed search patterns we found earlier in the cross-section. This exercise requires occupation-specific minimum wages to bind in all occupations. To validate this assumption, we use survey data to explore whether minimum wages are differentially binding across occupations. Table B.14 of Appendix B shows that higher-skill occupations exhibit higher average wages than lower-skill occupations but also higher minimum wages. The ratio between average hourly minimum wages and median hourly wages compares well across the different occupations, ranging from 6% to 18%. This suggests that the economic significance of the minimum wage is indeed comparable for the two occupational groups. 3.2 Empirical strategy Our empirical strategy exploits the frequent variation in minimum wages at the industry-by-occupation level provided by the CBAs. Each July and January, several industry-by-occupation cells (and, therefore, the vacancies within those cells) see their bargained wage floor adjusted (see Panel (c) of Figure 7). Other industry-by-occupation cells (and thus vacancies) see no change in their minimum wage, either because the contract is not adjusting wages in that adjustment window, or because some occupations do not have assigned minimum wages in certain sectoral contracts. This pattern of adjustments gives rise to natural treatment and control groups for which we can estimate DID models around the time of adjustment. The empirical strategy uses an industry-by-occupation cell as the unit of observation, for which we build a balanced panel to estimate standard event-study specifications. As a robustness check, we also estimate models using a within-vacancy design where the units of observation are vacancies that experience a minimum wage increase while being active (see Section 3.3). Our strategy may be interpreted as conservative for two reasons. First, as discussed above, we do not observe the exact minimum wage that is attached to each vacancy. Since we use the industry and occupation attached to the vacancy to measure exposure to minimum wage changes, which do not match one-to-one with the definitions in the CBAs, our estimates should be interpreted as reduced-form intend-to-treat (ITT) estimations, possibly inducing attenuation bias. Second, we assume that job applicants are aware of the timing of the minimum wage adjustments and, therefore, can update their applications after minimum wages are increased. Inattention to minimum wages should work against finding application effects and, therefore, should also exert downward bias in our estimations.29 28 Since CBAs change in a coordinated fashion, the measure of exposure does not depend on the choice of the statistic. 29 If the BJ platform included many vacancies targeting self-employed workers, who are not covered by minimum wages, this would be another source of possible downward bias. In practice, the fraction of such vacancies is negligible. For example, when searching for the pertinent expression “independiente” in job titles, only 110 (0.14%) mention it. 30 ILO Working Paper 136 Estimating equations In what follows, xit denotes the variable x of cell i in calendar time (month) t , where a cell is a 2-digit industry-by-1-digit occupation combination. Cells included in the balanced panel are cells for which we observe at least one posted vacancy during the whole period. In the data, we observe vacancies spanning 70 2-digit industries and 8 1-digit occupations. In terms of our unit of observation, we observe vacancies in 506 different cells (of a potential of 560). Our period consists on 108 months between October 2011 to September 2020, giving form to a total sample size of 54,648. Since treated cells potentially increase the minimum wage every six months, we implement a stacked event study as follows (Cengiz et al., 2019; Gardner, 2021; Baker et al., 2022; Dube et al., 2023). We define event periods ranging from three months before a minimum wage increase to two months after. This modeling decision means that event periods run either from October to March, or from April to September, such that preand post-event indicators are defined relative to January or July. Each event is indexed by e. We consider data from October 2011 to September 2020, which translates into 18 different event windows where a subset of the cells experiences a minimum wage increase. In each event, the subset of control cells is composed of cells with no minimum wage increase. Then, we estimate standard event specifications by allowing the cell fixed effects to vary by event. Since event periods do not overlap, time fixed effects automatically vary by event, and events are uniquely determined by calendar time, e(t). To add more flexibility, we also allow time fixed effects to vary by 1-digit industries. Formally, the estimating equation is given by: Y it is an outcome of interest for cell i in time t . Diτe(t) are event indicators, where τ denotes the distance from the event (in months) meaning that Diτe(t) is equal to one if cell i was treated τ months ago in event e(t). α ie(t) are cell-by-event fixed effects. Y j(i)t are month-by-1-digit industry fixed effects. X it are controls for the small share of minimum wage changes that occur in months different from January or July (see Panel (c) of Figure 7), whose effect is allowed to vary by event.30 Under the parallel trends assumption, βτ identifies causal effects from the minimum wage increase on Y it. As it is standard in event studies, β−1 is normalized to 0. Since minimum wage changes may be correlated within CBA across occupations, we cluster standard errors at the 2-digit industry level. To provide a quasi-experimental test for the cross-sectional directed search patterns documented in Section 3, we define Yit, our main outcome of interest, as the median number of applications received by vacancies of cell i posted in month t. We also consider a variation of equation (10) that interacts the event indicators with lowerand higher-skill occupation indicators. The sparsity of the balanced panel implies that Yit = 0 is a frequent outcome, so we estimate the equation in levels and then compute back-of-the-envelope estimates of the implied elasticity using external data on average minimum wage increases. As a complement to our directed search test, we also estimate effects of other outcomes at the cell-by-time level such as number of posted 30 Following Cengiz et al. (2019, 2021), Xit is computed as follows. Let tR be the month in which the rare minimum wage increase takes place. Then, define Earlyt = 1{t ϵ {tR3,tR2}}, Pret = 1{t = tR1}and Postt = 1{t ϵ {tR, tR+ 1,tR+ 2}}, and let Rarei be an indicator of cells that face rare minimum wage increases. Then Xit includes all the interactions between {Earlyt, Pret, Postt} × {Rarei} for each event separately. 31 ILO Working Paper 136 vacancies, openings, share of vacancies advertising non-wage amenities, and share of vacancies posting job requirements. To provide summary results, we also report estimates from standard pooled DID regressions: where Tie(t) is an indicator variable that takes value 1 if cell i is treated in event e(t), Postt is an indicator variable that takes value 1 if month t is equal to or larger than the treatment month (that is, if t corresponds to either January, February, March, July, August, or September), and all other variables are defined as in equation (10). The coefficient of interest in this specification is β. Table 11 presents descriptive statistics of the estimation sample. 47% of the observations exhibit at least one vacancy opening. The mean number of openings, including the zeros, is 4.62. The median and mean number of applications per opening is 35.33 and 43.89, respectively. 51% of the observations correspond to treated cell-by-event groups. Panels (b) and (c) break the statistics by lowand higher-skill occupations. The share of observations with at least one opening is remarkably similar across groups, although lower-skill vacancies usually exhibit more openings and more applications. Not surprisingly, lower-skill occupations are more likely to be treated than higher-skill occupations (60% versus 43%). 3.3 Application effects of the minimum wage Our main results use the median number of applications per opening at the cell-level as dependent variable. Figure 8 shows the estimated βτ coefficients of equation (10) with their corresponding 95% confidence intervals. Panel (a) shows the event study that pools all vacancies. The plot suggests that applications to vacancies in exposed cells increase after the minimum wage adjustments, although the increase is small and non-statistically significant at conventional levels. However, as shown in Panel (b), when interacting the event indicators with the lowand higher-skill occupation dummies, minimum wage increases tend to generate a significant increase in applications to exposed lower-skill vacancies with no effect on exposed higher-skill vacancies. Panel (c) plots the results from an analog triple difference regression, showing that the difference between lowerand higher-skill occupations is statistically significant, especially in the first month after the minimum wage adjustment. Table 12 shows the results of the pooled DID. Panel (a) shows results for the specification with no interactions, which estimates a noisy increase in 2.6 applications at the treated occupation-by-industry cell level, with an implied elasticity of 1.15 (Column (1)). Panel (b) shows the results for the specification with interactions for the occupational groups. The estimated effect for lower-skill vacancies is a significant average increase of 4.5 applications per cell, with an implied elasticity of 1.59. For higher-skill vacancies, the implied elasticity is only 0.02 (Column (1)). Two aspects of this result are worth discussing. First, the quasi-experimental exercise mirrors the cross-sectional finding of heterogeneous patterns of directed search across occupations. This is especially remarkable given the differences between the used variation, design, and set of vacancies. Second, the magnitude of the implied wage-applications elasticity in lower-skill vacancies is similar in magnitude, albeit on the lower end of the distribution, to the labor supply elasticity estimates documented in the empirical monopsony literature (Sokolova and Sorensen, 2021). This benchmark is reassuring given the reasons for which we expect attenuation bias in our regressions. 32 ILO Working Paper 136 Robustness checks and within-vacancy design Table 12 provides some robustness checks to our main estimates. For the sake of brevity, we focus the discussion on the specification including interactions by occupational group (Panel (b)). We also discuss below results from an alternative within-vacancy research design that yields similar results. Column (2) of Panel (b) shows that using mean instead of median applications per cell-by-time generates similar, albeit slightly smaller, estimates. Yet, the fact that the panel is sparse and the distribution of applications per vacancy is skewed suggests that the median may be a better-behaved measure at the occupation-by-industry-by-month level. Since the equation is estimated in levels, outliers may play a significant role in driving the results. Columns (5) and (6) show that excluding the event window-by-cell units for which the median number of applications per opening per cell in at least one month exceeded 750 attenuates the results, but the qualitative conclusions do not change. Finally, Columns (9) and (10) show that restricting the event window-by-cell units for which the cell had openings in at least two months barely changes the point estimates, although it decreases precision.31 As an additional robustness check, we estimate the application effects of the minimum wage using a different research design that exploits within-vacancy variation. We consider the sample of vacancies that are open when a minimum wage change potentially occurs (that is, either between June and July or between December and January) and estimate DID regressions at the vacancy level, comparing exposed and non-exposed vacancies before and after the policy change. This design allows us to refine the previous analysis as it controls for vacancy fixed effects and thereby captures time-invariant unobserved heterogeneity at the vacancy level. At the same time, since vacancies are usually open during only one month or less (see Table 1), we are not able to transparently assess the parallel trends assumption. The sample size, moreover, decreases because vacancies posted in other months are not included in the sample and because most applications to vacancies are made in the first few days of the vacancy period (see Figure B.8 of Appendix B).32 This pattern implies that we need to restrict to vacancies that have been open for a few days when the minimum wage change kicks in. Given these advantages and limitations, we see the within-vacancy design as a complement to the previous analysis, where we are interested in whether the main conclusions hold across the two empirical strategies. Our baseline sample of vacancies used in the within-vacancy design consists of the 2,129 vacancies (2.7% of the total sample) that were posted between June 25 and June 30 or between December 26 and December 31 in any of the years considered. Since we estimate the regression in levels, we exclude vacancies that receive more than 1,000 applications. Table B.15 of Appendix B shows descriptive statistics. With this sample, we estimate the following regression: where Y jt are the applications per opening to vacancy j in month t, T j is an indicator that takes value 1 if the vacancy j is treated, Postt is an indicator if month t is either January or July, αj are vacancy 31 We note that the implied elasticities are mechanically downward biased in the latter exercise because the exclusion of zeros distort the pre-event mean dependent variable by construction. 32 45% of applications happen in the first 2 days, 63% in the first 5 days, 72% in the first week. There are three explanations for this pattern. First, recent vacancies are more likely to appear first on the website. Second, when vacancies are filed, they may stop receiving applications. Third, applicants may opt for receiving emails with weekly updates of newly posted vacancies, which again may increase the salience of the recently posted ones. 33 ILO Working Paper 136 fixed effects, and γt are months (calendar time) fixed effects. As above, treatment status is determined based on the industry-by-occupation-by-adjustment window attached to the vacancy. Table B.16 of Appendix B shows the results. We estimate regression (12) for different sub-samples based on the days the vacancy was open before the potential policy change. Columns (1), (2), (3), and (4) show results for vacancies that were open for 3 days or less, 4 days or less, 5 days or less, and 6 days or less, respectively. We note two findings. First, while they are noisy, results imply that treated lower-skill vacancies receive an increase in applications after the minimum wage increase, with no corresponding effect in higher-skill vacancies. That is, the within-vacancy exercise supports the patterns documented so far. Second, the magnitude of the effect is decreasing in the days open before the policy change. The implied elasticity for lower-skill vacancies is 5.1, 2.6, 1.4, and 0.4 in the corresponding columns. This is consistent with vacancies being more salient when they are recently posted. Heterogeneity by applicant characteristics Table 13 presents results of equation (11) with occupation interactions but using applications from particular groups of applicants as the dependent variable. All groups exhibit larger respons - es for lower-skill vacancies relative to higher-skill vacancies. Consistent with the cross-sectional analysis, we find larger and more significant application responses to wages for male and older applicants. The implied wage-application elasticity to lower-skill vacancies is 2.3 for male applicants, relative to a non-significant estimate of 1 for female applicants. Likewise, the lower-skill elasticity for older applicants is 1.7, compared to a 1.4 estimate for younger applicants. The implied elasticities, however, are not different between employed and unemployed applicants, and are stronger for job seekers with no tertiary education, which contrasts from what was found in the cross-sectional analysis. This difference may be driven by the fact that, within occupation and industry, less educated applicants may be more attached to minimum wage jobs than highly educated applicants. 3.4 Additional results In the remainder of the section, we discuss results for complementary dependent variables. Vacancies and openings The positive effect of minimum wages on applications may come at the expense of a contraction in labor demand in terms of vacancies or openings. We test this hypothesis by estimating similar models as above using the total number of vacancies and openings per cell as the dependent variable. Figure 9 and Table 12 show that we do not find any detectable effect on vacancies and openings. This result suggests that the increase in applications may help firms buffer the increase in labor costs and/or that firms are adjusting other margins to pay for the increased minimum wage. Advertised non-wage amenities If providing amenities is costly for firms, advertised non-wage amenities could decrease after the minimum wage increase (Clemens, 2021). We test this hypothesis by estimating similar models as above using the share of vacancies that advertise non-wage amenities as the dependent variable. Figure B.6 and Table B.17 of Appendix B suggest the absence of negative responses on 34 ILO Working Paper 136 advertised amenities. The only advertised amenity that exhibits a non trivial negative implied elasticity is bonuses and commissions, however, the event studies suggest that the negative effect is possibly driven by differential pre-trends. Vacancy requirements Finally, firms could react to increased labor costs by becoming more selective in terms of education and skills requirements. Evidence of this narrative has been presented by Butschek (2021) and Clemens et al. (2021). We test for this hypothesis by estimating similar models as above using the share of vacancies that impose requirements as the dependent variable. Figure B.7 and Table B.18 of Appendix B suggest the absence of increases in education and skill requirements, although estimates are imprecise enough to make strong claims about these results. The only slightly significant positive estimate is an estimated increase in the share of higher-skill vacancies that require a college degree. 35 ILO Working Paper 136 XConclusions In this paper, we assess patterns of directed search in job applications, focusing on the role of posted wages and advertised non-wage amenities. Using rich data from a prominent online job board in Uruguay, we are able to provide a series of cross-sectional facts on job applications, which we then corroborate causally using plausibly exogenous minimum wage variation. First, we document substantial heterogeneity across applicants in the number of applications they send within an application spell, and find a large degree of diversification in terms of the occupations and industries of the vacancies they apply to within job seekers that send multiple applications. Second, we find robust evidence of directed search based on posted wages that is driven by vacancies attached to lower-skill occupations, with applications to vacancies attached to higher-skill occupations showing no responsiveness to posted wages. The directed search pattern is found to be stronger for male, employed, older, college-educated, and skilled job applicants. Finally, by applying text analysis to the job ads, we elicit advertised non-wage amenities and find that they play a key role in the application process. We find evidence of directed search based on amenities and show that applications to lower-skill vacancies are consistent with lexicographic job preferences where amenities affect applications only when wages are not posted. We also find substantial heterogeneity on the role of non-wage amenities by amenity, occupation, and applicant characteristics. The occupational heterogeneity in directed search is supported by a quasi-experimental exercise that uses minimum wage variation at the industry-by-occupation level to document positive application effects of minimum wage increases in lower-skill occupations. This exercise also suggests the absence of responses in the number of vacancies, openings, advertised amenities, or vacancy requirements after minimum wage increases. Our findings help inform several mechanisms behind the search-and-matching process in the labor market. They are consistent with models of directed search and suggest that industryand firm-wage differentials can be rationalized by the existence of rents rather than strong worker attachment to industries. They moreover unveil important occupational heterogeneities, which are consistent with the larger incidence of wage posting (rather than bargaining) in lower-skill occupations that has been documented in related literature. Based on our findings, several avenues of future research may be worth pursuing. First, it seems promising to explore the fundamental differences between occupations more deeply. While we conjecture that the differential incidence in wage posting and bargaining can explain these differences, further research is needed to depict a clearer picture of that pattern. Second, we have been able to exploit plausibly exogenous variation in wages, but additional causal analyses that rely on exogenous variation in amenities would further enhance the understanding of the job application process. Beyond their potential for informing economic theory, such analyses have practical value in that they shed light on how certain governmental interventions and firms’ recruitment strategies affect the applicant pool. 36 ILO Working Paper 136 References Adamczyk, Willian, Isaure Delaporte, and Verónica Escudero, “Job Amenities in Developing Countries: an analysis of online labour markets,” ILO Working Paper, Forthcoming. Alfonsi, Livia, Oriana Bandiera, Vittorio Bassi, Robin Burgess, Imran Rasul, Munshi Sulaiman, and Anna Vitali, “Tackling Youth Unemployment: Evidence From a Labor Market Experiment in Uganda,” Econometrica, 2020, 88 (6), 2369–2414. Allegretto, Sylvia and Michael Reich, “Are local minimum wages absorbed by price increases? Estimates from internet-based restaurant menus,” ILR Review, 2018, 71 (1), 35–63. Altmann, S., R. Mahlstedt, M. J. Rattenborg, and A. Sebald, “Which Occupations do Unemployed Workers Target? Insights from Online Job Search Profiles,” Working Paper, 2024. Arnold, David, Simon Quach, and Bledi Taska, “The impact of pay transparency in job postings on the labor market,” Working Paper, 2023. Ashenfelter, Orley and Štěpán Jurajda ˇ , “Minimum wages, wages, and price pass-through: The case of McDonald’s Restaurants,” Journal of Labor Economics, 2022, 40 (S1), S179—-S201. Atalay, Enghin, Phai Phongthiengtham, Sebastian Sotelo, and Daniel Tannenbaum, “The Evolution of Work in the United States,” American Economic Journal: Applied Economics, 2020, 12 (2), 1–34. Attanasio, Orazio, Adriana Kugler, and Costas Meghir, “Subsidizing Vocational Training for Disadvantaged Youth in Colombia: Evidence from a Randomized Trial,” American Economic Journal: Applied Economics, 2011, 3 (3), 188–220. Azar, José, Steven Berry, and Ioana Elena Marinescu, “Estimating labor market power,” Working Paper, 2022. Baker, Andrew, David F Larcker, and Charles CY Wang, “How Much Should We Trust Staggered Difference-In-Differences Estimates?,” Journal of Financial Economics, 2022. Banfi, Stefano and Benjamin Villena-Roldan, “Do high-wage jobs attract more applicants? Directed search evidence from the online labor market,” Journal of Labor Economics, 2019, 37 (3), 715–746. Barbanchon, Thomas Le, Roland Rathelot, and Alexandra Roulet, “Gender differences in job search: Trading off commute against wage,” The Quarterly Journal of Economics, 2021, 136 (1), 381–426. Bassier, Ihsaan, Alan Manning, and Barbara Petrongolo, “Vacancy duration and wages,” Working Paper, 2023. Batra, Honey, Amanda Michaud, and Simon Mongey, “Online Job Posts Contain Very Little Wage Information,” Working paper, 2023. 37 ILO Working Paper 136 Belot, Michèle, Philipp Kircher, and Paul Muller, “Providing Advice to Jobseekers at Low Cost: An Experimental Study on Online Advice,” The Review of Economic Studies, 2019, 86, 1411–1447. _ , _ , and _ , “How Wage Announcements Affect Job Search - A Field Experiment,” American Economic Journal: Macroeconomics, 2022, 14 (4), 1–67. Bó, Ernesto Dal, Frederico Finan, and Martín A Rossi, “Strengthening state capabilities: The role of financial incentives in the call to public service,” The Quarterly Journal of Economics, 2013, 128 (3), 1169–1218. Bonhomme, Stephane and Gregory Jolivet, “The pervasive absence of compensating differentials,” Journal of Applied Econometrics, 2009, 24 (5), 763–795. Burdett, Kenneth and Dale T Mortensen, “Wage differentials, employer size, and unemployment,” International Economic Review, 1998, pp. 257–273. Butschek, Sebastian, “Raising the bar: Minimum wages and employers’ hiring standards,” American Economic Journal: Economic Policy, 2021. Caldwell, S., I. Haegele, and J. Heining, “Bargaining and inequality in the labor market,” Working Paper, 2024. Caldwell, Sydnee and Nikolaj Harmon, “Outside options, bargaining, and wages: Evidence from coworker networks,” Working Paper, 2019. Card, David, “Who set your wage?,” American Economic Review, 2022, 112 (4), 1075–1090. _ , Jesse Rothstein, and Moises Yi, “Industry wage differentials: A firm-based approach,” Journal of Labor Economics, 2024, 42 (S1), 11–59. Cengiz, Doruk, Arindrajit Dube, Attila Lindner, and Ben Zipperer, “The effect of minimum wages on low-wage jobs,” The Quarterly Journal of Economics, 2019, 134 (3), 1405–1454. _ , _ , _ , and David Zentler-Munro, “Seeing Beyond the Trees: Using Machine Learning to Estimate the Impact of Minimum Wages on Labor Market Outcomes,” Journal of Labor Economics, 2021. Clemens, Jeffrey, “How do firms respond to minimum wage increases? Understanding the relevance of non-employment margins,” Journal of Economic Perspectives, 2021, 35 (1), 51–72. _ , Lisa B Kahn, and Jonathan Meer, “Dropouts need not apply? The minimum wage and skill upgrading,” Journal of Labor Economics, 2021, 39 (S1), S107–S149. Coviello, Decio, Erika Deserranno, and Nicola Persico, “Minimum wage and individual worker productivity: Evidence from a large US retailer,” Journal of Political Economy, 2022, 130 (9), 2227– 2500 Cullen, Zoë, “Is pay transparency good?,” Journal of Economic Perspectives, 2024, 38 (1), 153–180. Cullen, Zoe B, Shengwu Li, and Ricardo Perez-Truglia, “What’s my employee worth? The effects of salary benchmarking,” Working Paper, 2024. Deming, David and Lisa B. Kahn, “Skill requirements across firms and labor markets: Evidence from job postings for professionals,” Journal of Labor Economics, 2018, 36 (S1), S337–S369. 44 ILO Working Paper 136 Notes: This figure plots the statistic described in equation (1), the average number of “groups” individuals apply to in each quarter-by-year, as a function of the total number of applications made in the quarter-by-year to our final sample of posted vacancies (see Section 1 for details on the sample restrictions). “Groups” refer to 2-digit industries by 1-digit occupation cells (blue curve, which considers 504 categories), 2-digit industries (red curve, 70 categories), 1-digit industries (green curve, 14 categories), and 1-digit occupations (yellow curve, 8 categories). For readability, we censor the figure at 10 applications. Notes: This figure plots the share of applications made to vacancies attached to the same 1-digit occupation of the current employment as a function of the total number of applications made in the quarter-by-year to our final sample of posted vacancies 45 ILO Working Paper 136 (see Section 1 for details on the sample restrictions). By construction, this figure only considers applicants who are employed in the quarter-year of the application. As an example, around 15% of applications from job seekers employed as managers, who make 1 application in a given quarter-by-year, target managerial jobs (with the remaining applications targeting jobs in other 1-digit occupations); and around 8% of applications from job seekers employed as managers, who make 10 applications in a given quarter-by-year, target managerial jobs. Notes: This figure shows binned scatterplots and corresponding quadratic fits for the relationship between the log number of applications per vacancy and the log posted wage. The analysis considers all vacancies in our final sample that post a wage (see Section 1 for details on the sample restrictions). Panel (a) does not include controls. Panel (b) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Panels (c) and (d) augment Panel (b) specification by including 1-digit and 2-digit occupational fixed effects, respectively. 46 ILO Working Paper 136 Notes: This figure shows binned scatterplots and corresponding quadratic fits for the relationship between the log number of applications per vacancy and the log posted wage separately by occupation. The analysis considers all vacancies in our final sample that post a wage (see Section 1 for details on the sample restrictions). All plots exclude vacancies receiving more than 1,000 applications and include 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Panel (a) plots the aforementioned relationship for lower-skilled occupations (clerical support, services and sales, plant and machine operators, and elementary occupations). Panel (b) plots the aforementioned relationship for higher-skilled occupations (managers, professionals, technicians and associate professionals, and craft workers). Panel (c) plots the aforementioned relationship separately for the two aggregate occupational groups. 47 ILO Working Paper 136 Notes: This figure presents descriptive facts on the scheme of CBAs. Panel (a) shows the number of sectoral groups and related subgroups that determine the CBAs by year. Panel (b) shows the distribution of the number of minimum wages that are specified within each CBA by year. Panel (c) shows the monthly distribution of minimum wage adjustments, pooling all changes observed in our period. 48 ILO Working Paper 136 Notes: These figures plot the estimated βτ coefficients of equation (10) with their corresponding 95% confidence intervals using the median number of applications per vacancy within the sample of vacancies attached to the corresponding industry-by-occupation cell. Panel (a) pools all vacancies. Panel (b) considers interactions with indicators for lowerand higher-skill occupational groups. Panel (c) plots the corresponding triple difference, where the coefficient is interpreted as the difference between lowerand higher-skill vacancies. Regressions control for cell-by-event fixed effects, calendar month-by-1 digit-industry fixed effects, and minimum wage changes not occurring in January or July (see Panel (c) of Figure 7). Standard errors are clustered at the 2-digit industry level. Notes: These figures plot the estimated βτ coefficients of equation (10) with their corresponding 95% confidence intervals using different dependent variables. Coefficients are interacted with indicators for lowerand higher-skill occupational groups. Panel (a) uses the total number of posted vacancies as a dependent variable. Panel (b) uses the total number of openings as the dependent variable. Regressions control for cell-by-event fixed effects, calendar month-by-1 digit-industry fixed effects, and minimum wage changes not occurring in January or July (see Panel (c) of Figure 7). Standard errors are clustered at the 2-digit industry level. 49 ILO Working Paper 136 Notes: This table shows summary statistics. Panel (a) shows statistics for vacancies in our final sample (see Section 1 for details on the sample restrictions). Panel (b) shows statistics for applicants registered in the BJ platform. Panel (c) shows statistics for applicants at the time of application, only considering applications to our final sample of vacancies. In Panel (a), “vocational training” is defined as tertiary-level training, whereas the variable additionally captures lower levels of vocational training in Panel (b). 50 ILO Working Paper 136 Notes: This table shows summary statistics for the amenities advertised in our final sample of vacancies (see Section 1 for details on the sample restrictions). Advertised amenities were elicited following Adamczyk et al. (Forthcoming). The table details whereas vacancies advertise at least one amenity, the number of amenities advertised per vacancy, and provides information for each of the five individual amenities. Statistics are also shown separately between vacancies that post a wage and vacancies that do not post a wage. 51 ILO Working Paper 136 Notes: Panel (a) presents the estimated α coefficient of equation (2). Panel (b) presents the estimated (αLS, αHS) coefficients of equation (3). The dependent variable is the log number of applications, and the key regressor is the log posted wage, so coefficients are interpreted as cross-sectional elasticities. Column (1) shows results with no controls in Panel (a) and includes a control for the occupational group in Panel (b). Column (2) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively. Column (5) excludes the vacancies at the top 5% of the posted wage distribution. Column (6) only considers vacancies that are posted by firms that post at least 10 vacancies in the BJ platform and includes firm fixed effects. Compared to Column (2) of Table 2, the sample size for the specificiation without controls is marginally smaller due to a few vacancies receiving zero applications; results are robust to using a Poisson model (see Appendix Table B.2). Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 52 ILO Working Paper 136 Notes: This table presents the estimated (αLS, αHS) coefficients of equation (3). The dependent variable is the log number of applications made by applicants with the characteristic depicted in the column title, and the key regressor is the log posted wage, so coefficients are interpreted as cross-sectional elasticities. Panel (a) presents results by gender, employment status, and age. Panel (b) presents results by educational attainment (without tertiary education, vocational training, and college degree) and three categories of skills (cognitive, socio-emotional, and manual skills). Regressions exclude vacancies receiving more than 1,000 applications and include 2-digit industry fixed effects, year fixed effects, controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development), and 1-digit occupation fixed effects. Compared with Tables 2 and 3, discrepancies in sample sizes stem from vacancies receiving zero applications from the various groups; results are robust to using a Poisson model (see Appendix Table B.7). Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 53 ILO Working Paper 136 Notes: This table presents the estimated αa coefficients of equation (5). The dependent variable is the log posted wage, and the key regressors are indicators for advertised amenities, so coefficients are interpreted as cross-sectional semi-elasticities. Panel (a) presents results from regressions that include an indicator variable of advertising at least one amenity. Panel (b) presents results from regressions that include five indicators associated with individual amenities (bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Column (1) shows results with no controls. Column (2) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively. Column (5) excludes the vacancies at the top 5% of the posted wage distribution. Column (6) only considers vacancies that are posted by firms that post at least 10 vacancies in the BJ platform and includes firm fixed effects. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 60 ILO Working Paper 136 Notes: This table presents descriptive statistics of the estimation sample. The unit of observation is a 2-digit industry by 1-digit occupation cell by calendar month. Panel (a) shows summary statistics for all occupations combined. Panel (b) shows summary statistics for the lower-skilled occupational group. Panel (c) shows summary statistics for the higher-skilled occupational group. Notes: Panel (a) presents the estimated β coefficient of equation (11). Panel (b) presents the estimated β coefficients in a model that considers interactions with indicators for lowerand higher-skill occupational groups. Regressions include cell-by-event fixed effects, calendar month-by-1 digit-industry fixed effects, and minimum wage changes not occurring in January or July (see Panel (c) of Figure 7). The dependent variables, as depicted in the column titles, include (in levels) the median number of applications, the mean number of applications, the total number of vacancies, and the total number of openings. Reported elasticities 61 ILO Working Paper 136 are computed by dividing the β-coefficient by the pre-event average outcome within treated cells, normalized by the log change in minimum wage among treated cells. In each panel, Columns (1)-(4) present the main results, Columns (5)-(8) present results that exclude bin-by-event window observations where the median number of applications exceeded 750, and Columns (9)-(12) exclude bin-by-event window observations for which the outcome is 0 more than 4 months within the event window. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. Notes: This table presents the estimated β coefficient of equation (11) in a model that considers interactions with indicators for lower and higher-skill occupational groups. The dependent variable is the number of applications made by applicants with the characteristic depicted in the column title. Regressions include cell-by-event fixed effects, calendar month-by-1 digit-industry fixed effects, and minimum wage changes not occurring in January or July (see Panel (c) of Figure 7). Reported elasticities are computed by dividing the β-coefficient by the pre-event average outcome within treated cells, normalized by the log change in minimum wage among treated cells. Panel (a) presents results by gender, employment status, and age. Panel (b) presents results by educational attainment (without tertiary education, vocational training, and college degree) and three categories of skills (cognitive, socio-emotional, and manual skills). Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 62 ILO Working Paper 136 A. Methodology for Creating Variables from Free Text Entries This appendix provides an overview of the methodology used to create variables from free text entries, which is based on Adamczyk et al. (Forthcoming) and Escudero et al. (Forthcoming). We first discuss the creation of skills variables. We then discuss the creation of occupational codes. Finally, we discuss the elicitation of advertised non-wage amenities. Additional details can be found in the aforementioned papers. A.1 Skills All skill-related variables are based on the methodology developed in Escudero et al. (Forthcoming). Their approach seeks to provide a comprehensive representation of labor market dynamics across diverse contexts that goes beyond formal qualification measures by covering the skills demanded by employers in vacancies and highlighted by workers in their online profiles. The authors propose a taxonomy that groups skills into three broad categories: cognitive, socio-emotional, and manual skills. In turn, each category is decomposed into sub-categories, giving form to a total of fourteen subcategories. See Table A.1 for a description of each category and subcategory and the sources each category was derived from. The taxonomy is built upon existing literature from labor economics and psychology and has been expanded to adapt it to individual country contexts, with a particular focus on emerging and developing countries and online job board data. The starting point of the categorization is established taxonomies designed for classifying skills in online data within the United States, particularly Deming and Kahn (2018). Other sources used include Heckman and Kautz (2012), Kureková et al. (2016), and Deming and Noray (2020). The first extension is to include manual skills, which are often omitted in U.S.- centered analyses. Then, the second extension expands the conceptual foundations relating to cognitive and socio-emotional skills to facilitate a more comprehensive analysis of online data beyond individuals with high formal qualifications. To achieve these expansions, the taxonomy included additional keywords and expressions drawn from various studies (see Autor et al., 2003; Spitz-Oener, 2006; Almlund et al., 2011; Heckman and Kautz, 2012; Kureková et al., 2016; Hershbein and Kahn, 2018; Atalay et al., 2020; Deming and Noray, 2020), as well as the pilot exercise for O-NET Uruguay. To elicit the skill variables in the BJ data, the authors use a natural language processing (NLP) methodology that integrates pre-processing techniques with a rule-based classification approach, guided by the taxonomy and the specific list of keywords and phrases associated with each of the 14 subcategories. While some skills subcategories are closely linked, the keywords and expressions used to characterize them are distinct, allowing for the unique identification of skills in the data. In a second step, this dictionary is enlarged by including synonyms of the original words obtained through scraping a thesaurus website (www.wordreference.com) and manually checking the applicability of the retrieved synonyms. 63 ILO Working Paper 136 Notes: This table describes the skills subcategories presented in Table 1 of Escudero et al. (Forthcoming), based on the concrete keywords used in the taxonomy. ALM stands for Autor et al. (2003), APST for Atalay et al. (2020), DK for Deming and Kahn (2018), DN for Deming and Noray (2020), HK for Hershbein and Kahn (2018), HK for Heckman and Kautz (2012), KBHT for Kureková et al. (2016), and S-O for Spitz-Oener (2006). This procedure leads to a total of 741 distinct skills, based on the unique keywords and expressions. The skills-subcategory variables are then created using the unstructured text data present in both the vacancies posted by firms and the job spells of applicants available in their BJ profiles. From vacancy data, the authors elicit the skills that are demanded by the vacancy. From employment history data, the authors elicit the skills applicants report having. The open-text descriptions offer the most viable approach for creating skills variables, as they contain detailed information on skills for all vacancies (99.9%) and a majority of applicants’ job spells (68.5%). These opentext descriptions undergo a series of pre-processing steps using NLP techniques, including the translation of keywords and phrases from English to Spanish, tokenization, text normalization, lemmatization, n-gram creation in the skills taxonomy, and n-gram creation in the vacancy and 64 ILO Working Paper 136 applicants’ data. These processes are employed to reformat the text data into a structured format that facilitates the mapping with the skills dictionary. Finally, the skill variables are established by tallying keywords and phrases linked to each skill category and subcategory that are found within the text. A skill is considered present if at least one of the keywords/phrases from the dictionary is identified in the text. Additionally, we calculated the frequency of keyword occurrences for each skill and use this as an indicator of the degree of intensity with which a particular skill used. See Escudero et al. (Forthcoming) for additional details. A.2 Occupations The raw data provided by BJ only classifies vacancies and applicants’ job spells into ISCO-08 occupation codes for a limited subset of the data. This missing data problem prevents comprehensive analyses at the occupation level. To solve this problem, Escudero et al. (Forthcoming) employed a similar methodology as the one described aboved to elicit 1and 2-digit occupational codes for the full sample of vacancies and applicants’ job spells. To elicit the occupations posted vacancies seek to fill, the authors leveraged textual information from four open-text fields associated with each vacancy: job title, job description, required level of education, and hierarchical level of the position. To elicit the occupation associated with applicants’ job spells, the authors used the same information, except for job titles, which are not available as a separate entry. This data undergoes NLP procedures similar to the ones used for eliciting skills variables. The resulting text is then categorized into ISCO-08 codes through a three-step process. The first step is analogous to the rule-based model employed to create the skills variables. The authors employ a dictionary of keywords, selected based on the most frequently used words and phrases from the subset of both vacancies and applicants’ job spells already classified by BJ into ISCO-08 occupational codes. The dictionary used originates from the official ISCO-08 international classification. This exercise provided the set of rules used to classify the remaining job titles into occupational categories at the 2digit level. Additionally, the authors used information about the educational level to distinguish between levels 2 and 3, denoting individuals from the same field with either higher education (level 2) or any other education (level 3). Similarly, information about the hierarchical level is used to identify managers and directors, placing them in level 1 of the ISCO classification. To enhance the performance of the procedure, the authors introduced a machine learning algorithm (in the form of a predictive model) to assign codes to vacancies and job spells that were unclassified or for which the original BJ assignment significantly differs from the one that results from the algorithm. This process occurs in two steps. First, the model is trained using the already classified observations to assign 1-digit ISCO codes. Second, additional information from applicants and vacancies is incorporated into a second prediction model to refine the code assignment at the 2-digit level. Based on various tests and sensitivity analyses, the authors chose Gradient Boosting to code 1-digit and 2-digit occupations in the vacancy data, and Random Forest for the applicants’ data. As a result, 100% of vacancies have an assigned 1-digit occupation code, and 94.8% of them also have a 2-digit occupation code. For applicants, all job spells with a text description were classified at the 1-digit level, and 97.8% were also classified at the 2-digit level. See Escudero et al. (Forthcoming) for additional details. 65 ILO Working Paper 136 A.3 Amenities The methodology for identifying advertised amenities in unstructured vacancy data is akin to the one used for the skills variables, and it is based on the procedure outlined in Adamczyk et al. (Forthcoming). To begin, we developed a taxonomy of amenities using the related empirical literature as a starting point and then extending it to better suit the Uruguayan context and the nature of online job boards. As a first source, we follow Maestas et al. (2023), who provide a list of nine job attributes based on the results from the American Worker Conditions Survey (AWCS). The survey collects workers’ assessments of nine work characteristics: schedule flexibility, telecommuting opportunities, physical demands, pace of work, autonomy, paid time off, working with others, job-training opportunities, and impact on society. To broaden the scope of the categorization, we employ the comprehensive categorization proposed by Sockin (2024), which organizes non-wage amenities in 48 categories derived from the literature using a topic-modeling machine learning algorithm implemented in the text of amenities descriptions in U.S. employer-employee data. Table A.2 lists additional sources we use to refine the procedure for specific amenities. Given these categories, we then undertook three steps to broaden the scope of the categorization. First, we reorganized these categories to align with vacancy data. The literature primarily relies on U.S. workers’ reviews, but not all categories are pertinent to vacancy data because certain aspects of a job may not be appropriate to advertise in a posted vacancy. Second, we supplemented the list of keywords and expressions used in the literature to characterize different amenities, tailoring them to better fit the context of Uruguay. Third, we introduced an additional amenity category, “work equipment and allowances,” to reflect the post-pandemic reality and to incorporate attributes of manual work that may hold greater importance in Uruguay and other global south countries relative to the U.S. economy. We grouped these additional keywords into five broad categories, resulting in a total of 16 amenity subcategories. In some cases, we adjusted specific subcategories to ensure there was no overlap among the keywords and expressions assigned to each subcategory. The process yielded a final set of 659 words and expressions, comprising 357 original terms and 302 different versions of the same expressions (for multiword expressions). Table A.2 provides a list of these categories, along with their definitions and, where applicable, their sources in the literature. More details are available in Adamczyk et al. (Forthcoming). 66 ILO Working Paper 136 To apply this dictionary to the BJ vacancy data, both the terms in the dictionary and the free text information from the job advertisements need to be formatted appropriately. The process is similar to the one used for creating the skills variables, albeit with some modifications. These steps encompass keyword detection, tokenization (dividing the text into single units or tokens), normalization (removing capitalization and special characters), removing stop words (including exceptions for words included in the dictionary, such as ‘buen,’ ‘mucho,’ ‘gran,’ etc.), and 67 ILO Working Paper 136 Notes: See Adamczyk et al. (Forthcoming) for additional details. S stands for Sockin (2024), SS for Sockin and Sockin (2019), BE for Beckers et al. (2008), M for Maestas et al. (2023), SK for Simon and Kaestner (2004), G for Glassdoor (2015), L for Libert (2016), Q for Quinn (1974), WZ for Wasmer and Zenou (2002); LB for Le Barbanchon et al. (2020), MP for Mas and Pallais (2017), PPB for Park et al. (2021), BKS for Breza et al. (2017), HO for Holmlund (1983), HA for Hayward et al. (1989), NM for Neumark and McLaughlin (2012), FP for Filer and Petri (1988), HM for Hamermesh (1990), LLC for Lopes et al. (2014), AAZ for Athey et al. (2000), AP for Acemoglu and Pischke (1999), PR for Parent (1999), and BBB for Barron et al. (1999). lemmatization (associating different versions of a word, such as conjugated verb forms, with a common root word, like unconjugated verbs). Once the text describing vacancies and the keywords and expressions from the dictionary are in the same format, they can be matched using an NLP rule-based classification approach to identify amenities in the vacancy data. Importantly, this process accommodates variation in word order within expressions and allow matches with up to one external word in between the words from the dictionary expression. The algorithm then tallies the occurrences of words and expressions from the dictionary in the vacancy texts and aggregates them for each broader amenity category. To simplify the analysis, this number is transformed into an indicator variable for each amenity subcategory. The indicator takes the value of one if any of the keywords or expressions from that particular subcategory are identified in the job advert. Out of the 86,062 vacancies in the BJ data,33 50.6% were assigned at least one of the 16 amenities. While some vacancies list up to eight amenities, more than three-quarters of those with assigned amenities advertise only one or two. The most frequently matched subcategories are “human capital development” (22.6% of vacancies), “working in teams” (18.7%), and “work environment and impact on society” (17.9%). The lowest number of matches is found for “retirement contributions” (33 matches, or 0.04% of the observations) and “health insurance” (38 matches, 0.04% of the observations), possibly because these are legally mandated benefits that may not warrant explicit mention in the Uruguayan context. 33 The sample size mentioned here slightly deviates from sample sizes mentioned in Section 1. The filters to create our final analysis sample are not yet applied. Instead, the sample mentioned here excludes a few vacancies with blank or meaningless job text descriptions (see Atalay et al., 2020). In our main analysis, these vacancies are coded as having zero amenities. 68 ILO Working Paper 136 Notes: Analysis done on the base of 86,062 vacancies. See Adamczyk et al. (Forthcoming) for more details. A comprehensive list of the share of vacancies with the assigned amenity can be found in Table A.3. Regarding individual keywords, “trabajar en equipo (teamwork)”, which belongs to the working in teams category, is the most frequently matched (with a total of 12,579 matches). Typically, in each subcategory, a few keywords dominate the majority of matches, with other terms making smaller contributions. Figure A.1 displays word clouds for all amenity subcategories, where the size of a word corresponds to its share of matches within that subcategory. It is important to note that the use of keywords and expressions to create amenity variables underwent several rounds of manual verification to ensure that words and expressions were contextually accurate. This verification was manually conducted for a sample of vacancies for all words appearing at the top of the matches for each subcategory, as well as for a selection of other words deemed necessary by the authors of this study and Adamczyk et al. (Forthcoming). While the procedure was performed for the complete list of 16 amenities, in the analysis, we focus on the 5 amenities with the highest prevalence, namely, bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development. The rest of the amenities are found to be relatively infrequent and also, in some cases, offer additional interpretation challenges. 69 ILO Working Paper 136 Notes: Authors’ elaboration based on Adamczyk et al. (Forthcoming). The analysis is based on the full sample of 86,062 vacancies. The words displayed in the word clouds represent the original words used to define amenities. For the matching process, these original words were lemmatized to facilitate the matching. The inclusion of original words in the figure is for clarity and ease of understanding. 76 ILO Working Paper 136 Notes: This figure plots the distribution of the timing of applications to vacancies relative to the opening date. “Distance” refers to the days elapsed since the opening of the vacancy. Notes: This table shows summary statistics for the amenities advertised in our final sample of vacancies (see Section 2 for details on the sample restrictions). Advertised amenities were elicited following Adamczyk et al. (Forthcoming). The table details whereas 77 ILO Working Paper 136 vacancies advertise each individual amenity. This table considers the full list of amenities discussed in Appendix A. Statistics are also shown separately between vacancies that post a wage and vacancies that do not post a wage. Notes: Panel (a) presents the α coefficient of equation (2) estimated using a Poisson model. If β denotes the point estimate, the elasticity is recovered as exp(β) -1. The standard error is estimated using the Delta method. Panel (b) presents the estimated (αLS, αHS) coefficients of equation (3). The dependent variable is the log number of applications, and the key regressor is the log posted wage, so coefficients are interpreted as cross-sectional elasticities. Column (1) shows results with no controls in Panel (a) and includes a control for the occupational group in Panel (b). Column (2) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively. Column (5) excludes the vacancies at the top 5% of the posted wage distribution. Column (6) only considers vacancies that are posted by firms that post at least 10 vacancies in the BJ platform and includes firm fixed effects. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 78 ILO Working Paper 136 This table presents the estimated α coefficient of equation (2) using different definitions of posted wage. The dependent variable is the log number of applications, and the key regressor is the log posted wage, so coefficients are interpreted as cross-sectional elasticities. Panel (a) considers the midpoint of the salary range. Panel (b) considers the midpoint of the salary range, excluding vacancies whose range exceeds the 50% of the midpoint. Panel (c) considers the maximum of the salary range. Column (1) shows results with no controls. Column (2) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively. Column (5) excludes the vacancies at the top 5% of the posted wage distribution. Column (6) only considers vacancies that are posted by firms that post at least 10 vacancies in the BJ platform and includes firm fixed effects. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 79 ILO Working Paper 136 Panel (a) presents the estimated α coefficient of equation (2) for lower-skill occupations. Panel (b) presents the estimated α coefficient of equation (2) for higher-skill occupations. The dependent variable is the log number of applications, and the key regressor is the log posted wage, so coefficients are interpreted as cross-sectional elasticities. Within each panel, Columns (1)-(4) show results for individual occupations in regressions that exclude vacancies receiving more than 1,000 applications and include 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Within each panel, Columns (5)-(7) show results for the broad occupation groups. Column (5) does not include occupation fixed effects, while Columns (6) and (7) include 1-digit and 2-digit occupation fixed effects, respectively. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 80 ILO Working Paper 136 Notes: This table shows summary statistics for the presence of requirements (in terms of formal qualifications, foreign language, or skills) in our final sample of vacancies (see Section 1 for details on the sample restrictions). Panel (a) considers all vacancies. Panel (b) restricts to vacancies that post a wage. Within each panel, statistics are shown for all vacancies, vacancies attached to lower-skill occupations, and vacancies attached to higher-skill occupations. 81 ILO Working Paper 136 82 ILO Working Paper 136 Notes: This table presents the estimated (αLS, αHS) coefficients of equation (3). The dependent variable is the log number of applications, and the key regressor is the log posted wage, so coefficients are interpreted as cross-sectional elasticities. Panel (a) considers vacancies that post at least one formal qualification requirement (education and/or language). Panel (b) considers vacancies that post at least one skill requirement (cognitive, socio-emotional, and/or manual). Panels (c) and (d) present results without these requirements, respectively. Column (1) includes a control for the occupational group. Column (2) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively. Column (5) excludes the vacancies at the top 5% of the posted wage distribution. Column (6) only considers vacancies that are posted by firms that post at least 10 vacancies in the BJ platform and includes firm fixed effects. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 83 ILO Working Paper 136 Notes: This table presents the (αLS, αHS) coefficients of equation (3) estimated using a Poisson model. If β denotes the point estimate, the elasticity is recovered as exp(β) -1. The standard error is estimated using the Delta method. The dependent variable is the log number of applications made by applicants with the characteristic depicted in the column title, and the key regressor is the log posted wage, so coefficients are interpreted as cross-sectional elasticities. Panel (a) presents results by gender, employment status, and age. Panel (b) presents results by educational attainment (without tertiary education, vocational training, and college degree) and three categories of skills (cognitive, socio-emotional, and manual skills). Regressions exclude vacancies receiving more than 1,000 applications and include 2-digit industry fixed effects, year fixed effects, controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development), and 1-digit occupation fixed effects. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 84 ILO Working Paper 136 Notes: This table presents the estimated (αF, αE, αA, αV, αC, αCS, αSK, αMS) coefficients of equation (4). The dependent variable is the log posted wage of the application, and the key regressors are individual characteristics of the applicant. Column (1) shows results with no controls. Column (2) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively. Column (5) excludes the vacancies at the top 5% of the posted wage distribution. Column (6) only considers vacancies that are posted by firms that post at least 10 vacancies in the BJ platform and includes firm fixed effects. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 85 ILO Working Paper 136 Notes: This table presents the estimated ( αLS a , αHS a ) coefficients of equation (6). The dependent variable is the log posted wage, and the key regressors are indicators for advertised amenities, so coefficients are interpreted as cross-sectional semi-elasticities. Column (1) includes a control for the occupational group. Column (2) excludes vacancies receiving more than 1,000 applications and includes 2-digit industry fixed effects, year fixed effects, and controls for advertised amenities (indicators for bonuses and commissions, schedule flexibility, work environment/impact on society, working in teams, and human capital development). Columns (3) and (4) add 1-digit and 2-digit occupation fixed effects, respectively, in exchange for the occupational group indicator. Column (5) excludes the vacancies at the top 5% of the posted wage distribution. Column (6) only considers vacancies that are posted by firms that post at least 10 vacancies in the BJ platform and includes firm fixed effects. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 92 ILO Working Paper 136 Notes: This table presents the estimated β coefficient of equation (11) in a model that considers interactions with indicators for lower and higher-skill occupational groups. Regressions include cell-by-event fixed effects, calendar month-by-1 digit-industry fixed effects, and minimum wage changes not occurring in January or July (see Panel (c) of Figure 7). The dependent variables, as depicted in the column titles, include (in levels) the share of vacancies requiring vocational training, the share of vacancies requiring a college degree, the share of vacancies requiring foreign language knowledge, the share of vacancies requiring cognitive skills, the share of vacancies requiring socio-emotional skills, and the share of vacancies requiring manual skills. Reported elasticities are computed by dividing the β-coefficient by the pre-event average outcome within treated cells, normalized by the log change in minimum wage among treated cells. Standard errors (reported in parentheses) are clustered at the 2-digit industry level. 93 ILO Working Paper 136 Appendix Bibliography Acemoglu, Daron and J. S. Pischke, “The structure of wages and investment in general training,” Journal of Political Economy, 1999, 107 (3), 539–572. Adamczyk, Willian, Isaure Delaporte, and Ver´onica Escudero, “Job Amenities in Developing Countries: an analysis of online labour markets,” ILO Working Paper, Forthcoming. Almlund, Mathilde, Angela Lee Duckworth, James Heckman, and Tim Kautz, “Chapter 1 - Personality Psychology and Economics,” in Eric A. Hanushek, Stephen Machin, and Ludger Woessmann, eds., Handbook of the Economics of Education, Vol. 4, Elsevier, 2011, pp. 1–181. Atalay, Enghin, Phai Phongthiengtham, Sebastian Sotelo, and Daniel Tannenbaum, “The Evolution of Work in the United States,” American Economic Journal: Applied Economics, 2020, 12 (2), 1–34. Athey, Susan, Christopher Avery, and Peter Zemsky, “Mentoring and Diversity,” American Economic Review, 2000, 90 (4), 765–786. Autor, David, Frank Levy, and Richard Murnane, “The skill content of recent technological change: An empirical exploration,” The Quarterly Journal of Economics, 2003, 118 (4), 1279–1333. Barbanchon, Thomas Le, Roland Rathelot, and Alexandra Roulet, “Gender Differences in Job Search: Trading off Commute against Wage,” The Quarterly Journal of Economics, 2020, 136 (1), 381–426. Barron, John M., Mark C. Berger, and Dan A. Black, “Do Workers Pay for On-The-Job Training?,” The Journal of Human Resources, 1999, 34 (2), 235. Beckers, Debby G.J., Dimitri Van Der Linden, Peter G.W. Smulders, Michiel A.J. Kompier, Toon W. Taris, and Sabine A.E. Geurts, “Voluntary or involuntary? Control over overtime and rewards for overtime in relation to fatigue and work satisfaction,” Work & Stress, 2008, 22 (1), 33–50. Breza, Emily, Supreet Kaur, and Yogita Shamdasani, “The morale effects of pay inequality,” The Quarterly Journal of Economics, 2017, 133 (2), 611–663. Deming, David and Kadeem Noray, “Earnings Dynamics, Changing Job Skills, and STEM Careers,” Quarterly Journal of Economics, 2020, 135 (8), 1965–2005. and Lisa B. Kahn, “Skill requirements across firms and labor markets: Evidence from job postings for professionals,” Journal of Labor Economics, 2018, 36 (S1), S337–S369. Escudero, Verónica, Hannah Liepmann, and Ana Podjanin, “Using Online Vacancy and Job Applicants’ Data to Study Skills Dynamics,” Research in Labor Economics, Forthcoming. Filer, Randall K. and Peter A. Petri, “A Job-Characteristics Theory of Retirement,” The Review of Economics and Statistics, 1988, 70 (1), 123. Glassdoor, “4 in 5 Employees Want Benefits or Perks More Than a Pay Raise; Glassdoor Employment Confidence Survey (Q3 2015) - Glassdoor US,” 2015. 94 ILO Working Paper 136 Hamermesh, Daniel S., “Shirking or Productive Schmoozing: Wages and the Allocation of Time at Work,” Industrial and Labor Relations Review, 1990, 43 (3), 121S. Hayward, Mark D., William R. Grady, Melissa A. Hardy, and David Sommers, “Occupational influences on retirement, disability, and death,” Demography, 1989, 26 (3), 393–409. Heckman, James and Tim Kautz, “Hard evidence on soft skills,” Labour Economics, 2012, 19 (4), 451–464. Hershbein, Bratt and Lisa B. Kahn, “Do recessions accelerate routine-biased technological change? Evidence from vacancy postings,” American Economic Review, 2018, 108 (7), 1737–72. Holmlund, Bertil, “Payroll Taxes and Wage Inflation: The Swedish Experience,” The Scandinavian Journal of Economics, 1983, 85 (1), 1. Kureková, Lucia Mýtna, Miroslav Beblavý, Corina Haita, and Anna-Elisabeth Thum, “Employers’ skill preferences across Europe: between cognitive and non-cognitive skills,” Journal of Education and Work, 2016, 29 (6), 662–687. Libert, Kelsey, “Employee Benefits Study: The Cost and Value of Employee Perks,” 2016. Lopes, Helena, Sérgio Lagoa, and Teresa Calapez, “Work autonomy, work pressure, and job satisfaction: An analysis of European Union countries,” The Economic and Labour Relations Review, 2014, 25 (2), 306–326. Maestas, Nicole, Kathleen J Mullen, David Powell, Till Von Wachter, and Jeffrey B Wenger, “The value of working conditions in the United States and the implications for the structure of wages,” American Economic Review, 2023, 113 (7), 2007–2047. Mas, Alexandre and Amanda Pallais, “Valuing alternative work arrangements,” American Economic Review, 2017, 107 (12), 3722–59. Neumark, David and Joanne Song McLaughlin, “Barriers to Later Retirement: Increases in the Full Retirement Age, Age Discrimination, and the Physical Challenges of Work,” Working Paper, 2012. Parent, Daniel, “Wages and Mobility: The Impact of Employer-Provided Training,” Journal of Labor Economics, 1999, 17 (2), 298–317. Park, R. Jisung, Nora Pankratz, and A. Patrick Behrer, “Temperature, Workplace Safety, and Labor Market Inequality,” IZA Working Paper, 2021, (14560). Quinn, Robert P., Job Satisfaction: Is There a Trend?, University of Michigan Library, 1974. Simon, Kosali Ilayperuma and Robert Kaestner, “Do Minimum Wages Affect Non-Wage Job Attributes? Evidence on Fringe Benefits,” ILR Review, 2004, 58 (1), 52–70. Sockin, Jason, “Show Me the Amenity: Are Higher-Paying Firms Better All Around?,” Working Paper, 2024. and Michael Sockin, “A pay scale of their own: Gender differences in variable pay,” Working Paper, 2019, (3512598). 95 ILO Working Paper 136 Spitz-Oener, Alexandra, “Technical Change, Job Tasks, and Rising Educational Demands: Looking outside the Wage Structure,” Journal of Labor Economics, 2006, 24 (2), 235–270. Wasmer, Etienne and Yves Zenou, “Does City Structure Affect Job Search and Welfare?,” Journal of Urban Economics, 2002, 51 (3), 515–541. 96 ILO Working Paper 136 Acknowledgements We thank Marcelo Bergolo, Sydnee Caldwell, David Card, Rodrigo Ceni, Zoe Cullen, Jennifer Hunt, Patrick Kline, Elva López Mourelo, and seminar participants at the Columbia Junio MicroMacro Labor Conference, the 6th IDSC of IZA Workshop (“Matching Workers and Jobs Online”), Princeton University, the 8th RDW Conference (ILO), UC Berkeley, and Universidad de la República de Montevideo for very helpful comments and suggestions. We especially thank Marcelo Bergolo, Paula Carrasco, Rodrigo Ceni, Nicolás Grau, and Cecilia Parada for generously sharing raw data on digitized collective bargaining agreements, and Willian Boschetti Adamczyk for contributing to the analysis of non-wage amenities. The responsibility for opinions expressed in this article rests solely with its authors, and publication does not constitute an endorsement by the International Labour Office of the opinions expressed in it. Other usual disclaimers apply. Research Department (RESEARCH) International Labour Organization Route des Morillons 4 1211 Geneva 22 Switzerland T +41 22 799 6530 resear[email protected] www.ilo.org/research XAdvancing social justice, promoting decent work The International Labour Organization is the United Nations agency for the world of work. We bring together governments, employers and workers to improve the working lives of all people, driving a human-centred approach to the future of work through employment creation, rights at work, social protection and social dialogue. I S B N 9789220413920 9HSTCMA*ebdjca+ 9 789220 413920 Contact details