The Age Twist in Employers’ Gender Requests

A. The Job-Specificity of Employers’ Gender Requests

Since firms typically place multiple ads on Zhaopin, Kuhn and Shen (2013) were able to ask what share of firms post at least one ad requesting a man and at least one requesting a woman. Among firms who placed more than 50 Zhaopin job ads, 38.6 percent placed at least one ad that favors men and one ad that favors women. In XMZYJS, XMRC, and Computrabajo, these shares were 93.4, 96.2, and 89.2 percent, respectively (see Online Appendix 2 for details). These patterns support the idea that gender-typing of jobs within firms plays a major role in the age twist.

A second way to assess the relative role of firms and jobs in accounting for patterns of gender requests is to decompose the variance of advertised gender requests across ads into components that are associated with occupations, with firms, with differences in the way firms assign genders to occupations, and within-job cell (firm*occupation cell) variation. The occupation controls used in this analysis are the categories that each of the job boards uses to classify ads on their site, which number 58 in XMZYJS, 36 in XMRC, 39 in Zhaopin, and 14 in Computrabajo, respectively. Online Appendix 2 shows that less than 40 percent of the variance in advertised gender requests can be accounted for by occupation and firm fixed effects in all four of our data sets. In addition, all data sets exhibit a large amount of within-job-cell (firm*occupation) variation in advertised gender requests, ranging from about one-third of the total in Zhaopin and XMRC to more than 70 percent in Computrabajo.

Taken together, the above evidence suggests that models based on firm-level tastes for men relative to women are likely to be of little help in understanding patterns in gender profiling. A more appropriate class of models would view firms as having a portfolio of jobs, some of which they view as best suited to men and others to women. The importance of the type of work done, rather than the identity of the firm doing the hiring, for gender profiling is dramatically confirmed in Section IV below, where we introduce detailed job descriptions derived from job titles in addition to firm fixed effects.

B. The Negative Skill–Targeting Relationship

As noted, Kuhn and Shen (2013) found that Zhaopin jobs requiring higher skill levels were less likely to be gender-targeted and less likely to be age-targeted. Using education requirements as our main skill measure, Figure 1 of this paper confirms that this relationship holds for both age and gender targeting within all four of our data sets. By disaggregating gender-targeted ads into those requesting women versus men, Figure 1 also shows that in most cases, explicit requests for both women and men decline with job skill requirements. This shows that the skill–targeting relationship is not an artifact of a tendency for employers to prefer one gender in skilled jobs and the other in unskilled jobs.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

The Negative Skill–Targeting Relationship—Share of Ads That Are Age- and Gender-Targeted, by Desired Education

Notes: <HS, HS, SC, and C denote less than high school, high school, some college, and college or more, respectively. CT refers to the Computrabajo data.

Kuhn and Shen (2013) also showed that the skill–targeting relationship was robust to different measures of skill (education, experience, and/or wages) and to regression controls for occupation effects, firm effects, and their interaction. This is confirmed for our three new data sets in Online Appendix 3. The relationship persists when we use required experience or the offered wage as our measure of skill requirements and in regression specifications that are as similar as possible to those of Kuhn and Shen (2013) in the three new data sets, including occupation and occupation*firm controls.

C. What Explains the Negative Skill–Targeting Relationship?

As we have noted, one interpretation of the skill–targeting relationship, suggested by the Kuhn and Shen (2013) model, is that it is a direct consequence of job skill requirements. Since higher skill levels “raise the stakes”—making it more important to identify the best job candidate—simple screening models predict that firms should search more broadly as jobs’ skill demands rise. Kuhn and Shen (2013), however, also identify other possible causes of the skill–targeting relationship, including skill-based variation in stigma against employers posting targeted ads, in application processing costs, in the expected number of applicants per position, and in the idiosyncratic variance across jobs in their relative suitability for men versus women (σ_v in the model). In the latter case, the argument is that while men and women may be quite different in their suitability for unskilled jobs (with systematic differences in fine manual dexterity versus gross motor strength, for example), such gender differences are less important in skilled or mental tasks.

In Online Appendix 4, we replicate a series of tests designed by Kuhn and Shen (2013) to distinguish among the above explanations and conclude, as Kuhn and Shen (2013) do, that uncontrolled variation in stigma, in application processing costs, and σ_v are not likely explanations, but that a decline in the number of applications per job (that is, increasing labor market tightness) with skill could help account for the negative skill–targeting relationship, provided that this tightness pattern applies in all four of our data sets and is robust to all our regression controls.

Summing up, evidence from our three new data sets shows that the negative skill–targeting relationship first identified by Kuhn and Shen (2013) extends to the use of both gender and age targeting, in four different samples of job ads serving workers of very different skill levels in two emerging economies. The relationship is robust to controls for the type of work done (firm*occupation fixed effects) and holds for three different measures of skill requirements: education, experience, and the level of the posted wage. Additional tests show that the relationship is most likely a direct result of skill requirements (or of systematically tighter labor markets for skilled jobs in all four of our data sets) than of other unmeasured factors. A potential implication is that the rapid, ongoing skill upgrading in many emerging economies may itself lead to a voluntary reduction in employers’ use of coarse demographic screens like age and gender to prejudge the suitability of potential job applicants.

A. Documenting the Age Twist

The age twist in employers’ gender requests in all four of our data sets is shown in Figure 2. While the share of ads making a gender request (the sum of the male and female bars) does not vary much with the worker’s age, in all cases the direction of firms’ gender requests shifts strongly away from women and towards men as the desired age of the worker rises. Consider, for example, the XMZYJS data. If a firm on this job board is looking for a worker younger than 25, the odds that it is seeking a woman (versus a man) are about 1.4 to one (46/32). On the other hand, among ads requesting workers older than 35, requests for men outnumber requests for women by more than 2.5 to one (58/22 = 2.64). This pattern is even more dramatic in our other three data sets. In XMRC ads requesting women outnumber ads requesting men by more than four to one when firms are seeking workers younger than 25, while ads requesting men outnumber ads requesting women by five to one when firms are seeking workers older than 35. In Zhaopin ads run more than five to one in favor of women for workers younger than 25 and more than four to one in favor of men for workers 35 and older. Finally, in Computrabajo these ratios are two to one favoring women among those younger than 25 and 2.5 to one favoring men among those older than 35.¹¹

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

The Age Twist—Share of Ads Requesting Women and Men, by Desired Age

Notes: The share of job ads in each of the four age groups (younger than 25, 25–29, 30–34, and 35+) are 0.30, 0.38, 0.26, and 0.06 in XMZYJS; 0.16, 0.41, 0.30, and 0.12 in XMRC; 0.12, 0.30, 0.33, and 0.25 in Zhaopin; and 0.07, 0.33, 0.40, and 0.19 in Computrabajo.

What explains this striking pattern? As a first step in understanding it, we ask to what extent these correlations could be associated with differences in the types of skills employers demand from workers of different ages or by the possibility that men and women may work at different firms as they age. To that end, we run regressions of the “direction” of employers’ gender requests on various features of the job ad, including occupation indicators. Following Kuhn and Shen (2013), our measure of the direction of employers’ gender requests is given by (P^M – P^F), where P^M (P^F) is an indicator for whether the ad requests a man (woman). Thus, (P^M – P^F) equals −1 when the job requests women, zero when the ad is not gender-targeted, and 1 when the ad requests men. Kuhn and Shen (2013) show that under reasonable conditions this approach reveals the determinants of firms’ underlying assessments of men’s and women’s relative desirability for the job being advertised.¹²

To interpret the effect of occupation and other controls on the estimated age twist, let β^U > 0 (the unadjusted age twist) be the coefficient on an age indicator in a regression without the above controls: (1) and let β^A (the regression-adjusted twist) be the coefficient from a regression that controls for whether the job is, say, in Occupation 1, X¹: (2)

Using the omitted-variable bias formula, controlling for Occupation 1 helps account for the age twist (β^A < β^U) iff: (3) where D¹ is estimated from the auxiliary regression: (4)

Thus, there are two types of occupations that can help explain the age twist: (i) occupations for which firms disproportionately request men (γ¹ > 0) and request older workers more frequently than younger workers (D¹ > 0) (henceforth “growing male” occupations) and (ii) occupations for which firms disproportionately request women (γ¹ < 0), but women are less frequently requested as workers age (D¹ < 0) (henceforth “shrinking female” occupations). Examples might be management and low-level administrative positions, respectively.

Estimates of the above “gender preference” regressions for each of our data sets are presented in Tables 3–6.¹³ Aside from the age controls, the specifications are identical to the regressions used to investigate the skill–targeting relationship in Online Appendix 3. Column 1 of each table has no controls, thus reproducing the results in Figure 2. Column 2 adds controls for the job’s education requirement, firm characteristics (size and ownership type), the number of job openings advertised, and occupation fixed effects.¹⁴ Column 3 replaces the firm characteristics by a full set of firm fixed effects, and Column 4 interacts the firm and occupation fixed effects.

Overall, there are two main messages from Tables 3–6. First, occupation and firm controls account for a substantial share (ranging from 32 to 66 percent depending on the data set and which age category is compared to ages 18–24) in XMZYJS and Zhaopin, but for much more modest shares in the remaining data sets. Indeed, the attenuating effect of these variables is essentially absent in Computrabajo, which has the coarsest occupation categories. Second, the estimated twist remains large in magnitude and highly statistically significant in all four data sets, even in the presence of the occupation*firm controls in Column 4. Essentially, these Column 4 regressions compare two job ads for the same occupation in the same firm (say, sales at Dell Computer) requesting workers of different ages. They show that when an older worker is requested, the firm is more likely to request a man relative to a woman compared to when a young worker is requested.

B. Understanding the Age Twist: Job Titles and the Role of Gendered Tasks

The attenuation of the estimated age twist when occupation controls are introduced in Tables 3–5 suggests that more detailed age-related differences in the type of work employers offer to men and women of different ages might account for more of the twist. For example, suppose that employers specifically prefer young women over other age–sex groups for customer-contact jobs. Or suppose that employers prefer their truck drivers (who are overwhelmingly male) to be older because young men are seen as unreliable. A large or influential set of jobs like this—where employer, customer, or coworker tastes or expected productivity depend on age–sex interactions—could explain the age twist.

To explore this idea further, we are able to access much more detailed information about job duties—the job title—in one of our data sets: XMRC. Job titles are the main heading in every job ad. They provide a brief description of the job, and can run up to 12 words in XMRC. For example, here is a random sample of ten job titles on the XMRC website: front desk administration assistant, project engineer, quality control, shift leader, customer service maintenance specialist, administration, ME product engineer Xiamen, experienced two-dimension designer, customer service engineer, and front desk clerk. Job titles appear to provide considerably more relevant information about the type of work than even the most granular standardized occupational classification systems. For example, Marinescu and Wolthoff (2019) found that job titles on Careerbuilder.com were much more predictive of advertised wages than six-digit SOC codes, and were essential controls for identifying the effect of advertised wages on the number and quality of applications an ad received.

Job titles are highly predictive of employers’ gender requests. This is illustrated in Table 7, which focuses on just the first word of the title and lists the 25 “most female” words, the 25 “most male” words, and the 25 words that are least likely to be explicitly gendered.¹⁵ Some first words that occur very frequently in the data (like front desk, clerk, and teller) are strongly associated with explicit requests for female employees, while others (security, electronic, and driver) are strongly associated with requests for men.¹⁶ This suggests that job titles might help us understand not only employers’ gender requests, but how they interact with their age requests.

To create a sample of XMRC job titles for analysis, we translated and cleaned all job titles that appeared more than once in the XMRC data set. This gave us a data set of 9,884 job ads (relative to 17,021 in Table 4), which appears to be representative— running the Table 4 regressions on this subsample yielded almost identical results. To quantify the ability of these new job title variables to account for the age twist, it is helpful to have a scalar measure of the magnitude of the twist (as opposed to the four age categories used in Figure 2 and Tables 3–6). Accordingly, in Table 8 we measure the job’s requested age by a single dummy variable for whether the requested age of the worker is 30 or older.¹⁷

Columns 1 and 2 of Table 8 reproduce the specification in Columns 1 and 2 of Table 4 for our sample of translated titles with a scalar age twist indicator. They show, once again, that adding broad occupation controls attenuates the estimated twist somewhat: from 44.2 to 37.2 percentage points. Columns 3–6 successively add fixed effects for the first k = 1,…,4 words of the raw job title, while Column 7 has fixed effects for every title, regardless of length. Column 8 adds a full set of firm fixed effects to the regression. According to the table, controlling for just the first word of the job title attenuates the estimated age twist considerably more (0.372 – 0.257 = 11.5 percentage points) than a full set of occupation controls (0.442 – 0.372 = 7.0 percentage points). Adding the second and third words attenuates the estimated twist further (to 17.0 percentage points), after which neither additional words nor a full set of firm fixed effects changes the age coefficient substantially. In the most saturated specification of Table 8 (Column 8), the estimated age twist equals 0.155. Thus, in total, a differing mix of detailed types of work and (to a lesser extent) a different mix of firms hiring old versus young workers, together can account for ([0.442 – 0.155]/0.422 =) 65 percent of the age twist.

What specific types of jobs account for employers’ declining explicit interest in female workers and increasing requests for male workers as workers age? To answer this question, we manually inspected large numbers of titles and noted that their most frequently occurring words fell into four main categories: words denoting the job’s rank or skill level; words suggesting frequent contact with customers, clients, or business associates; miscellaneous words describing the worker, job, or firm (but not the job duties); and (most commonly) words describing the type of work done (duties). We then coded these words as follows.

C. Decomposing the Age Twist: Which Job Characteristics Matter Most?

While successively adding covariates to a regression sheds some light on which ones are most important in accounting for an empirical regularity, the estimated importance of different covariates is not invariant to the order in which they are added to the regression. Thus, while Table 9 suggests that employers’ demand for young, female Assistants and Contacters plays an important role in the age twist, in this section we use an approach recently extended by Gelbach (2016) to uniquely decompose the unadjusted age twist in our data (0.442) into components associated with various covariate groups. In addition to order invariance, Gelbach’s approach also has the advantage of yielding consistent estimates of economically and econometrically meaningful population parameters that are based on the omitted variable bias formula, and of producing consistent estimates of the standard errors of the covariate groups’ contributions. Such estimates are often absent in previous decomposition exercises.

In our context, as suggested by Equations 1–4, the ability of a job characteristic, X, to account for the age twist depends both on its effects on employers’ gender preferences (g) and on its correlation with employers’ age requests (D). Generalizing Equation 2 to the case of multiple regressors, (5) where x_i¹ and x_i² are vectors of variables in two different categories (for example, contact variables and top-100 words), the omitted-variable bias formula allows us to partition the explained portion of the age twist into components related to each of the two variable groups: (6) where each element of d^J (D^kJ) is estimated from a different auxiliary regression: (7)

Figure 3 uses Equation 6 to decompose the explained twist of 0.2609 in Table 9 into the contributions of 11 distinct variable groups, showing 95 percent confidence bands.²¹ Of these, the largest single contributor (0.0761) is the occurrence of a helping word in the job title. Five other categories of variables also contribute substantial amounts, specifically the detailed duty words from the job title (0.0667), the presence of a contact word in the job title (0.0484), an indication in the title that the job contains some managerial duties (0.0268), differences in occupation mix between jobs requesting older versus younger workers (0.0249), and firm fixed effects (0.0142). The remaining five variable categories—education requested, miscellaneous worker words, miscellaneous job words, number of vacancies advertised, and expertise words—contribute negligible amounts. Importantly, the large role of job titles relative to the quite muted role of a full set of firm fixed effects underscores the Kuhn and Shen (2013) job-specificity result: explicit gender requests are very much about the type of work that is done and much less about which firm is doing the hiring.²²

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

“Explained” Components of the Age Twist, XMRC

Notes: Contributions are measured in percentage points and are computed using Equation 6 from Columns 1 and 6 in Table 9. The total explained age twist equals 0.2609. Lines indicate 95 percent confidence intervals.

Figures 4–6 provide additional detail on the contributions of individual variables within the five quantitatively important variable categories in Figure 3. According to Figure 4, just three occupations—administration, construction, and sales (0.0123 + 0.0091 + 0.0034 = 0.0248)—can account for the entire occupation mix effect (though some smaller countervailing effects exist). Of these, construction is a “growing male” occupation (occupations that constitute a larger share of explicitly gendered job ads for older workers than younger workers), while administration and sales are influential female occupations that shrink in size with age.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

Sources of the Occupation Mix Component of the Age Twist, XMRC

Notes: This figure decomposes the occupation mix component in Figure 3 (0.0249 percentage points in total) into four types of occupations. Individual occupations making the ten (three) largest absolute contributions are: growing male occupations (γ > 0, d > 0)—construction (0.0091), manufacturing (0.0034), logistics (0.0025), management (0.0010); shrinking male occupations (γ > 0, d < 0)—software (−0.0023); growing female occupations (γ < 0, d > 0)—home service (−0.0022), finance/accounting (−0.0014); shrinking female occupations (γ < 0, d < 0)—administration (0.0123), sales (0.0038), international trade (0.0020).

Figure 5a shows that growing requests (with worker age) for higher management positions (which are disproportionately male) account for the majority of the contribution of the Managers variables to explain the twist. Figure 5b shows that (0.0473/0.0761 =) 62 percent of the Assistants effect comes from explicit requests for female clerks, with smaller contributions from assistants and secretaries (Figure 6). According to Figure 5c, the largest single contributor to the Contacters effect on the twist comes from explicit requests for young women in front desk positions, with customer service and teller titles also contributing substantial shares.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5

Sources of the Manager, Assistant, and Contacter Components of the Age Twist

Notes: This figure allocates the contributions of the three components (0.0268, 0.0761, and 0.0484 percentage points, respectively) to the more detailed subcomponents shown.

Finally, turning to the duty words, Figure 6 shows that growing male title words contribute considerably more to the age twist than shrinking female words (though both are counteracted by substantial contributions from shrinking male and growing female words). To provide some context for these effects, the growing male job title words making the ten largest contributions to the age twist are driver (0.0282), electronic (0.0130), security (0.0123), engineer (0.0044), chef (0.0044), plumber (0.0027), molding (0.0025), fitter (0.0022), equipment (0.0021), and carpenter (0.0019). Together, these words suggest that employers’ disproportionate requests for older versus younger men in a variety of semiskilled occupations also play a quantitatively significant role in the twist.²³

• Download figure
• Open in new tab
• Download powerpoint

Figure 6

Sources of the “Detailed Job Duties” Component of the Age Twist, XMRC

Notes: This figure decomposes the “detailed job duties” component in Figure 3 (0.0667 percentage points in total) into four types of words. Individual words making the 20 (five) largest absolute contributions are: growing male words (g > 0, D > 0)—driver (0.0282), electronic (0.0130), security (0.0123), engineer (0.0044), chef (0.0044), plumber (0.0027), molding (0.0025), fitter (0.0022), equipment (0.0021); shrinking male words (g > 0, D < 0)—network (−0.0024), warehousing (−0.0023); growing female words (g < 0, D > 0)—accounting (−0.0116), finance (−0.0065), division (−0.0056), cleaning (−0.0051); shrinking female words (g < 0, D < 0)— statistician (0.0052), omitted group (0.0028), business (0.0025), international trade (0.0022), cuisine (0.0021).

Summarizing all the above information, Figure 7 apportions the explained age twist to its main sources. Most strikingly, 53 percent of the explained age twist is attributable to just three sets of factors: a tendency to request older men in managerial positions (9 percent); a tendency to request young women in four customer contact jobs: front desk, customer service, teller, and cashier (17 percent); and a tendency to request young women in three helping positions: clerk, assistant and secretary (27 percent). That said, a disproportionate number of requests for older men in the following occupations (and job titles) also play a role: the construction occupation, and driver, electronic, and security as job title words.

• Download figure
• Open in new tab
• Download powerpoint

Figure 7

Sources of the Explained Portion of the Age Twist, XMRC

Notes: The explained portion of the age twist (comparing Columns 1 and 8 of Table 8) is 44.20 −15.48 = 28.72 percentage points, or 65 percent of the unadjusted twist. This figure summarizes the main factors contributing to this explained twist. “Uncoded aspects” refers to the difference between fixed effects for every unique title (Table 8) and Table 9, which codes ranks and individual words in the titles. Arrows show the top three absolute occupation contributions, the top five absolute title–words contributions, and the most important helping and contact job words. F and M denote “female” and “male” occupations or title words, (γ < 0 and γ > 0, respectively). Title words make positive contributions to explaining the twist when γD > 0. Thus “growing” (with age) (D > 0) male title words like “driver” and “electronic” help account for the twist, while growing female title words like accounting and finance actually contribute negatively to the twist.

Importantly, the results in Table 9 and Figures 3–7 are robust to all the different ways we have explored to encode the rich information contained in job titles. For example, very little changes when we control for many fewer title words than the full 515 of them.²⁴ The intuition is simply that infrequently appearing job duties can only have a limited explanatory effect on a bounded outcome variable; thus, the main results are driven by specific title words that are already listed in Figure 7. Online Appendix 5 shows that the results are also robust to replacing dummies for occurrence of individual duty words by fixed effects for 880 mutually exclusive combinations of all those words.²⁵ Results are also very similar when we model requests for men (P^M) and requests for women (P^F) as separate dependent variables instead of as a difference. Here, we find that the share of ads explicitly requesting men rises with age and the share of ads explicitly requesting women falls with age. Both of these effects attenuate as controls are added, in much the same way that their difference attenuates. Indeed, the estimated effects of age on (P^M – P^F) from this procedure are arithmetically identical to those reported in Tables 8 and 9.

D. Exploring the Contacters and Assistants Effects: The Role of Beauty

The previous section estimated that employers’ requests for young women in contact and helping positions can account for a substantial share of the age twist [about (0.0761 + 0.0484)/(0.442 – 0.1811) = 47.7 percent of the explained component and 0.47 × 59 = 28 percent of the total]. But why do employers prefer young women in these roles? Motivated by literatures on beauty and customer discrimination in labor markets, here we ask whether a desire for physical attractiveness plays a role.²⁶ Unfortunately, our analysis is limited by the fact that we do not have data on explicit beauty requests in the XMRC data.²⁷ Still, we can draw some inferences using data on employers’ explicit requests for beauty in the Zhaopin and Computrabajo data sets. As documented in Table 1, employers request attractive workers in 4.0 percent of ads on the Computrabajo site and in 7.7 percent on Zhaopin. These beauty indicators were derived by searching the text of the detailed job descriptions for phrases requesting a visually attractive candidate, then compiling a list of the most common forms of such requests. Online Appendix 1 lists the ten most common requests in both data sets.²⁸ In the Computrabajo data we have access to a second indicator of firms’ interests in the candidate’s attractiveness: 10.9 percent of Computrabajo ads instructed applicants to include a photo with their application.²⁹

Using the above data we are able to establish the following facts concerning the role of beauty in the age twist.

Fact 1. Employers are more likely to request beauty, or to request a photograph, when they are requesting young women than any other demographic group.

For the Zhaopin data, Fact 1 is illustrated in Panel A of Figure 8. There, close to one-half (45.8 percent) of job ads aimed at women under the age of 30 explicitly request that the applicant be physically attractive. This number falls to 25.5 percent among women older than 30, 9.8 percent for men under 30, and to 4.8 percent for men older than 30. Similar but less dramatic trends are present for beauty and photo requests in the Computrabajo data (Panels B and C). Together, these patterns confirm that employers’ requests for beauty covary strongly with their age and gender requests and do so in a direction that could help account for the age twist.

• Download figure
• Open in new tab
• Download powerpoint

Figure 8

Beauty Requests by Age and Gender

Fact 2. In the data sets where we have information on explicit employer requests for beauty, those requests account for 7–15 percent of the age twist.

To estimate the extent to which employers’ explicit beauty requests can account for the estimated age twist, we added a control for such requests to the gender request regressions in Tables 5 and 6. The results, reported in Online Appendix Table A5.3, show that beauty (and photo) requests strongly predict explicit requests for women and that controlling for beauty attenuates the estimated age twist by 15.25 percent in Zhaopin and by 7.50 percent in Computrabajo.

A final piece of evidence concerning the role of beauty in the age twist comes from which occupations have the largest age twists.

Fact 3. Looking across all four of our data sets, the age twist is especially high in administration, customer service, management, and sales occupations. The age twist is also significantly higher in occupations with higher-than-average beauty demands.

Online Appendix 5 shows the size of the age twist in each occupation in all four of our data sets. While the twist is positive in almost all occupations in all data sets, four occupations stand out as having consistently high twists: administration, customer service, management, and sales.³⁰ The disproportionate role of these occupations is consistent with the substantial role of contact, helping, and managerial job title words in the XMRC data because these title words appear very frequently in those occupations.³¹ Online Appendix 5 (Table A5.4) also reports the results of cross-occupation regressions in the Zhaopin and Computrabajo data sets. In both data sets, occupations with a higher demand for beauty exhibit significantly larger age twists (p < 0.01 in Zhaopin and p < 0.10 in Computrabajo).

In sum, data on employers’ explicit requests for physically attractive applicants that are available in two of our four data sets suggest that requests for young, female beauty, especially in certain key occupations like administration and customer service, account for 7–15 percent of the age twist. It is worth noting, however, that according to the Kuhn and Shen (2013) model, firms will not always explicitly request worker characteristics they believe are correlated with productivity; requests for any characteristic—including beauty—should only be made when employers’ priors about its productivity effects are sufficiently extreme or precise. If that is the case, the true contribution of beauty preferences to the age twist, especially in contacting and helping jobs, may be considerably greater than is indicated by explicit beauty requests only.

E. Exploring the Within-Title Twist: Marriage, Parenthood, and Labor Force Attachment

In Section IV.C, we established that 65 percent of the age twist can be explained by job titles and other ad characteristics. What factors might account for the remaining 35 percent, which occurs within detailed job titles? Having accounted for job characteristics, some promising candidates are worker characteristics like marriage, parenthood, and any related changes in labor force attachment, turnover, and work intensity. Marriage and parenthood change substantially between the ages of 18 and 45 and might have differential effects on men’s and women’s attractiveness to employers.³² To explore these ideas, this section begins by establishing two key facts about employers’ explicit requests for marital status, and about the timing of the age twist in China and Mexico. These facts suggest a connection between the twist and parenthood. We conclude by asking whether this apparent connection might be linked to changes in men’s and women’s relative labor force attachment, turnover, and work intensity.

Fact 4. Explicit employer requests for workers’ marital status are rare in China and Mexico. When they occur, however, their direction is highly correlated with explicit gender requests—employers prefer men to be married, and women to be single.

Information on employers’ requests for marital status are only available in our Computrabajo data, where only 3.4 percent of ads explicitly indicate a preferred marital status.³³ That said, the pattern of these requests strongly suggests that employers prefer married men and single women. Specifically, among ads requesting men, Figure 9 shows that requests for married applicants outweigh requests for single applicants by a ratio of about six to one. In ads requesting women, this pattern is reversed, with requests for single applicants outweighing those for married applicants by more than four to one.

• Download figure
• Open in new tab
• Download powerpoint

Figure 9

Share of Ads Requesting Single and Married Applicants, by Requested Gender (Computrabajo Data)

Fact 4 suggests that the movement of men and women into marriage during the early career has the potential to account for some of the age twist. Quantitatively, however, adding controls for marital status requests into the Table 6 regressions has a negligible effect on the estimated twist, since these requests are so rare. As in the case of beauty requests, however, this may understate the true contribution of employers’ marital status preferences to the twist if only the most extreme marital status preferences are actually written into job ads.

A second clue regarding the role of employers’ marital status and parenthood preferences comes from looking at the precise timing of the age twist, taking advantage of the fact that Chinese women’s entry into marriage and motherhood is highly concentrated into a specific age range.

Fact 5. In urban China, the age twist in advertised gender requests is highly concentrated between the ages of 26 and 30. This corresponds quite closely to the ages at which most urban Chinese women are having their first (and only) child.

The three dashed lines in Figure 10A and 10B show the share of gendered job ads that request women in each of our three Chinese data sets in two-year age bins. For comparability, each data set has been reweighted to represent the education distribution of working-age urban Chinese women in the 2005 1 percent census microdata file. After reweighting, the timing of the age twist is quite comparable across our three Chinese data sets (especially the two less skilled ones, XMRC and XMZYJS). Remarkably, most of the age twist is concentrated in the four years between ages 26 and 30, with about 70 percent of gendered job ads for 26-year-olds requesting women compared to about 35 percent at age 30.

• Download figure
• Open in new tab
• Download powerpoint

Figure 10

Timing of Marriage and Motherhood Relative to the Age Twist

Notes: (A) Solid lines show the share of urban Chinese women who are single at that age or who will still be single two years later (from the 2005 census microdata sample). (B) Solid lines show the share of urban Chinese women who are childless or who will still be childless two years later (from the 2005 census microdata file). (C) Solid lines show the share of urban Mexican women who are childless at that age, or who will still be childless two years later (from the 2010 census microdata file). Dashed lines in all parts of the figure show the share of gendered job ads that request women on each job board, reweighted to match the education distribution of urban women in the country.

Panel A of Figure 10 compares the timing of this rapid shift in the gender mix of job ads to the ages at which urban Chinese women move into marriage in the 2005 census sample described above. In Panel A, the right-most, darker solid line shows the share of urban Chinese women who are single in two-year age bins. Interestingly, urban Chinese women’s movement into marriage is also highly concentrated into a narrow age range, with the fraction single falling from greater than 90 to less than 30 percent between the ages of 20 and 26. This range, however, is earlier than the period when the mix of gendered job ads shifts most strongly away from women (26–30). Thus, the timing of the age twist does not line up well with a marriage bar (Goldin 1991) or employer-aversion-to-married-women hypothesis. Chinese employers are still disproportionately requesting female employees (at around 70 percent of gendered ads in XMRC and Zhaopin) at age 26, when a substantial majority of women are already married.³⁴ This timing problem becomes even more severe if we expect employers to avoid not just married women but women who are likely to marry shortly after being hired; leading the “share single” curve by two years (the lighter solid curve on the left) shows the share of women who, if hired today, will still be single two years from now. To ensure they hire such women, firms would need to start shifting their ads away from women starting even earlier—at age 20.

Panel B of Figure 10, on the other hand, compares the age twist in urban China to the ages at which urban Chinese women tend to have their first child. Here, the curves match up better. The composition of gendered job ads shifts rapidly away from requesting women during exactly the ages when Chinese women are becoming mothers. While hardly conclusive, this suggests that a reluctance to hire mothers relative to fathers may also contribute to the age twist.³⁵ Notably, however, the share of ads requesting women does not recover at ages when children are in school. Given the timing of first births and China’s one-child policy, Online Appendix Figure A5.7 shows that the share of urban Chinese women without a child under the age of six begins to increase rapidly after age 34. If care for preschool-age children explains the age twist, we should see increased employer interest in hiring women at these ages, which is not the case.

For comparison, Panel C of Figure 10 compares the timing of Mexico’s age twist to the share of women who are childless in that country. These trends are consistent with each other in the sense that both time trends are much more gradual in Mexico than in China, but the fact that Mexican women’s fertility is more evenly spread over the life cycle does not offer as sharp a test in Mexico as it does in China.

Is the age twist in employers’ gender requests between ages 26 and 30 in urban China accompanied by similarly pronounced changes in labor force activity at those ages? This question is explored in Figure 11 and Online Appendix 6. Figure 11a, for China, shows that both men’s and women’s employment rates rise during their 20s as many are completing school. Perhaps surprisingly, however, both genders’ employment rates are remarkably stable between the ages of 26 and 40. Men’s employment rate holds steady at about 90 percent and women’s at 70 percent. Thus, the age twist in employers’ advertised requests for female employees does not coincide with an exodus of Chinese mothers from the labor force.³⁶ Interestingly, the same is true in Mexico. According to Figure 11b, the sizable gender gap in employment rates is surprisingly constant after age 26, which is about when Mexico’s age twist in job ads begins (Figure 10c).

• Download figure
• Open in new tab
• Download powerpoint

Figure 11

Employed Share of the Population, by Age

Sources: Authors’ calculations from the Chinese 2005 1 percent census and Mexican 2010 census microdata samples.

If not labor force withdrawal, could the within-job-title age twist be connected to changes in other aspects of labor force activity, such as turnover and work intensity? Figures A6.2–A6.6 consider this question for China, beginning with the issue of turnover. Our best available measure of turnover in the Chinese census is the share of employed workers with a contract of two years duration or fewer (suggesting a lower level of past and prospective job attachment than a longer contract); this is shown in Online Appendix Figure A6.2.. While women are somewhat more likely to have such short-term contracts, and while employed men and women both transition out of such short-term contracts at a substantial rate as they age, the gender gap in the prevalence of such contracts is remarkably constant over the work life.³⁷

If men become more committed to their jobs when they become fathers, and women less, we should expect to see sizable shifts in work hours around the age of parenthood. Online Appendix Figure A6.3 shows that this is not the case either. While employed, private-sector men report working one to two more hours per week than women do, both men and women work long hours—between 49 and 53 hours per week—and the gender–hours gap is remarkably small and constant between the ages of 26 and 42. Related, Figure A6.4 shows that part-time work is extremely rare among both men and women in urban China, and Figure A6.5 shows that large and relatively constant shares of both men and women work more than 50 hours per week.³⁸

Having ruled out large changes in labor force attachment and work hours, what other factors might account for the within-title age twist around the age that Chinese workers are getting married? One possibility is an effect of motherhood and fatherhood per se on workers’ productivity. For example, men could be becoming more reliable workers as they become fathers, or simply as they enter their 30s. Women could devote less energy to their jobs at the same ages. Indeed, even though employed women’s work hours do not change around the age of motherhood, time-use data show a large (nine hours per week) increase in Chinese women’s unpaid work time around age 25 (Dong and An 2015, Table 4), which could affect employers’ propensities to request female job applicants. Alternatively, customer and coworker tastes could play a role. Even within job titles, coworkers and customers may expect to interact with a certain type of person (attractive young women selling clothes, older men supervising younger coworkers), and employers may be reluctant to violate those expectations. Such factors would also help explain the fact that employers’ very low propensity to request women on all our job boards does not recover long after their children have reached school age.

We conclude this section by discussing one potential explanation of the age twist we cannot address directly. This explanation results from the fact that—like all vacancybased data—our data only capture recruiting that occurs via the formal process of posting a job ad. Thus, while we are confident that our two Xiamen-based data sets provide a reasonably complete picture of the formal vacancies in a large Chinese city, our data cannot rule out the effects of a gender-specific change in the way firms recruit workers as they age. In this regard, it seems likely that the share of workers who are recruited via formal methods, that is, by firms posting ads and soliciting applicants, declines as workers age. As workers gain labor market experience, they are likely to acquire more connections and to develop a reputation and to substitute towards less formal job search methods that exploit these advantages. To explain the age twist, however, women’s use of informal job search networks would need to rise more rapidly with age than men’s. While this seems unlikely—if anything, we might expect women to acquire fewer “old boy” connections as they age—we are not aware of any data that can conclusively rule it out.³⁹