Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=

# Screening Women Out? Pay Transparency in Job Postings

Authors:   
Discussant: None  
Video: https://www.youtube.com/watch?v=mC4ESywpIDc&t=16814s


## Talk (04:40:14 – 05:31:25)

[04:40:14] >> Awesome. >> Thanks so much for having me on the program. I'm Aean. I am a PhD student at the LSE for three more days and then I'll join Harvard. This was my job market paper titled screening women out  
[04:40:28] pay transparency in job postings. This project is motivated by the well-known fact that almost everywhere in the world women's education and employment rates are on the rise. But it remains the case  
[04:40:40] that women are clustered in low-wage and low productivity firms. And this pattern of sorting across firms can now account for up to three4s of the gender pay gap.  
[04:40:52] So essentially we've made some progress on women's labor force participation but not so much their allocation. And to figure out what's the right policy lever for addressing that allocation challenge. It's important to disentangle  
[04:41:06] whether these patterns of sorting reflect women's preferences for the low wage firms or barriers to accessing the high wage firms. The preference story is something that we may naturally turn to.  
[04:41:18] Many papers have shown us that women's employers are not just lowpaying, they're also more familyfriendly, located closer to where women live, have flexible or shorter hours, and so on. So then it's tempting to read that literature as saying perhaps we don't  
[04:41:33] have a market failure, but rather women just really like some amenities and are willing to pay for them. My contention with that reading of the literature is that in order for women's preferences to explain the observed wage gaps, it would  
[04:41:47] have to be that their willingness to pay for these amenities is quite large, about 30% of wages. But most lab estimates of such preferences are orders of magnitude smaller. For example, at  
[04:41:59] most 8% for remote work. So while the preference story is really important, it just doesn't add up to explain the scale of the problem that we have at hand. In that setting, then uh this paper  
[04:42:12] proposes a new fact that can help us make progress on understanding how women end up at low-wage firms. And that fact is that at the time the workers are searching for jobs, they don't necessarily know the salaries of the  
[04:42:25] jobs. So on the x-axis here I have the share of job ads that do not contain salary information and on the y- axis are a bunch of high and low-income countries for which I could access this data. What you can see at the top is  
[04:42:38] that across the spectrum of development Chile, US, Ethiopia, China, Slovenia, Germany about 80% of job ads don't say anything about salaries. In Pakistan where this study is going to be based about 55% of job ads don't say anything  
[04:42:52] about salaries. Now why this fact matters for the question they raised is that if we think women are trading wages for amenities it's possible they don't know the price they end up paying for those amenities. That's something that  
[04:43:04] the existing types of uh papers and research designs that we have uh abstract from. So you'll either have lab designs that are aiming to get at women's willingness to pay for certain  
[04:43:17] amenities. to detect true uh preferences they have to abstract from information or search frictions of the kind that this graph uh talks about. In doing that they get at the preferences um but erase  
[04:43:31] the role that frictions play outside the lab in shaping worker search behavior.  
[04:43:35] Then you have match employee data which contain equilibrium matches. Those are a product of both preferences and frictions. But to know whether it's one or the other, uh we don't just need to know whether men and women do different  
[04:43:48] jobs, but whether they apply to different jobs for a given choice set.  
[04:43:53] And this latter thing is what I'm going to get at in this paper because I'm going to observe workers as they search for jobs within the information environment that shapes that search. To do that, I'm going to partner with Pakistan's largest online job search  
[04:44:06] platform. This platform is bigger than its next five competitors put together.  
[04:44:11] It gets about 200 job ads a day from a range of occupations and industries that you could imagine searching for workers online. And a key feature of this platform is that it's going to allow me to observe the salaries even when the  
[04:44:25] workers do not because the platform always collects this data from the firms but then allows them to optionally hide it from the job seekers. With that, I'm going to first um characterize some equilibrium facts uh through a series of  
[04:44:39] different types of data. I'm first going to use platform data itself which is going to have about 29 million job applications, 62,000 firms, millions of workers over a six-year period. I'm going to supplement that with firm and worker surveys and then do my own  
[04:44:53] discrete choice lab experiment to get at preferences over amenities, wages, and information about both of them.  
[04:45:00] I'm then going to use that to write down a simple theoretical framework that shows that uncertainty around wages depresses the utility of the wage component of the job and instead magnifies the utility of the amenity  
[04:45:14] component. And that magnification is why you can have a world where uh while gender differences and preferences can be small, they can drive large gender gaps in applications.  
[04:45:26] I'm then going to test the predictions of this framework with the help of a large-scale field experiment. The experiment uh is implemented in partnership with this platform and it will mandate pay transparency in the treatment group. So by that I mean that firms will be told you have to report  
[04:45:41] the minimum and maximum salary for this randomly selected job to be posted online. But in the control group whether or not they want to reveal this information is optional. The experiment runs for about 7 months. It affects about 20,000 jobs, 8,900 firms and at  
[04:45:55] least 310,000 workers were exposed to it because that's how many people sent applications over this period.  
[04:46:01] Underlying the experiment is also going to be a saturation design at the occupation industry level that gets at potential spillovers.  
[04:46:10] Okay, so let me tell you what I find. Uh first through the equilibrium facts I'm going to characterize four descriptive facts that come together to screen women out of high-paying jobs. The first is the demand side one which is that the  
[04:46:24] large firms in particular pay well. That's not new and has been documented in the literature. But what's new in this paper is that those well-paying firms are more likely to hide their salaries in job ads. There's a number of reasons they report to me in surveys for  
[04:46:38] doing that which I'll get at towards the end of the talk. But crucially, it's not theoretically uh obvious that that would be their optimal strategy. And so workers do not realize this. I find that on the workers end both men and women  
[04:46:52] sort on wages when the wages are accessible but absent wage information.  
[04:46:57] They don't know this correlation and they're not able to figure out the wages based on the non-wage information of that job. So there is then this genderneutral policy window for pay transparency to inform the beliefs of workers and redirect their research. But  
[04:47:12] what makes this a gendered story is that these firms that are well-paying not just hide that fact but on the margin also offer less flexibility particularly in work hours and location. And that is something that women moderately care  
[04:47:24] more about than men. So together we have this perfect storm where wages are hidden exactly where they could be high enough to compensate workers for the lack of flexibility and that's why women  
[04:47:37] self- select out of these jobs. That's where my experiment then comes in. I find that mandating paid transparency uh does not change the wage or amenity landscape of the market but increases women's applications to these high wage  
[04:47:52] firms by about 95%. Men's applications also increase because they also learn something new but only by 59%. And that not just closes but in fact reverses the gender gap in directed search towards these firms. On the firm side, I find  
[04:48:07] that the same large firms that are more likely to hide their salaries in job ads after the experiment uh if they were more exposed to the experiment which which is at the job level become 30% more likely to start voluntarily  
[04:48:20] disclosing these salaries which suggests that at least for some of the firms they had not internalized the benefits of paid transparency until the experiment forced them to experience them.  
[04:48:30] What I want you to take away from all of this at the end of the day is that uh it's very common to think that women prefer amenities over wages and that drives the gender pay gap. But what I want to show in this paper is that it's rather the lack of information about wages that amplifies the role of  
[04:48:45] amenities. Okay. So here's going to be the road map of the talk. I'm first going to show you some facts, then use them to write down a theoretical framework, motivate the design of the experiment, talk mainly about the workers results, the mechanisms that underly them. This  
[04:49:00] mostly a paper about workers, but at the end come back to why do firms hide salaries and why do I think that they changed their behavior after the experiment.  
[04:49:11] Okay, so happy to take questions at any point. Um, what I'm going to do throughout the talk is think of good firms where we would like to see more women as simply large firms because that's a catchall proxy  
[04:49:24] for pay and productivity and so on. In the next slide, I'm going to show to you how these large firms differ from small firms in terms of their wages, their wage disclosure policies, and then their amenities.  
[04:49:36] The first bar here is saying that large firms pay on average 15% more than small firms holding all else about the job that I can observe in the ad constant.  
[04:49:46] But the second bar is saying that these same high-paying firms are also 16 percentage points or about 30% more likely to hide the salary in the job.  
[04:49:56] What that means is that in equilibrium for the same job title and a vast variety of other characteristics that are constant, higher salaries are more likely to be hidden. That's not something that workers realize. So when I survey  
[04:50:10] workers on this platform, only 11% of both men and women think that hidden salaries are higher than posted ones. So once again, there's clearly a search friction that binds to both genders, but at the moment, it's not clear why that  
[04:50:24] would have gendered implications when I reveal salaries. What makes this a gendered story is the fact that these large firms are not just paying more and hiding more, they're also allowing for less flexibility. So specifically, um  
[04:50:38] the likelihood, holding again job characteristics constant, that you're allowed flexible hours is about 10 percentage points lower. the likelihood that you can work from home is about six percentage points or 35% lower. So what I'm going to do now for the rest of the  
[04:50:53] paper is take these facts about uh firm level wages and amenities is constant and think about the implications that has for worker search with and without pay transparency. Yes.  
[04:51:06] >> No amenities that flexibility if they don't know wages.  
[04:51:14] So um the specific amenities that I'm going to focus on and that I see descriptively drive most of my results are very commonly reported in all job ads in this context. They're systematically collected by the platform and those are going to be work hours and where the work is like whether it is at  
[04:51:28] home or in the office. Those are things that are commonly reported generally but in this context for sure they're collected. Um it is an object that could be moved by my intervention but I find that forcing them to reveal salaries  
[04:51:41] does not change their amenity supply. equilibrium. I'm seeing you know where I  
[04:51:54] see your workers maybe being at the partial equilibrium exercise characteristics. Yes, it's possible for example that what I the reason I think of them as equilibrium is I observe them at baseline they remain the same even  
[04:52:07] during my experiment but it's possible that this is a short-term experiment and then if this were a law firms may change this behavior >> yeah just the salaries that we mentioned the firms that didn't support them where's that data coming from  
[04:52:21] >> so yeah that's what I was mentioning earlier which is that the platform collects this data from the firms but then they can optionally hide it from the jobseker So I always observe it and crucially I find that um when they are forced to reveal the salaries they reveal are very similar to the ones that  
[04:52:36] they report to the platform. So it's quite credible that this information is correct. Yes.  
[04:52:46] I mean is do the firm have the firms need to commit to that level of salaries in six months.  
[04:52:52] >> Yeah. So they don't commitment is not necessary but I think it's quite clear both from the surveys and also from the fact that some firms go back to their previous behavior of hiding salaries and I should make clear the default is that the salary is shown so firms have to opt out of that default. The fact that they  
[04:53:07] take that action suggests that there are costs. The costs they report to me in the survey are workers just anchor to especially the max and then it makes bargaining quite hard. The other cost is screening. When we advertise very good salaries, a lot of people who are  
[04:53:20] irrelevant will start applying. So they they do think that there are costs. Yes.  
[04:53:24] So just very quick question to understand generally if you think about incomplete like the worker having complete information about the salary itself you can think of like expensive margin I see the salary or not but then like I'm thinking about context of  
[04:53:39] platforms in Nigeria and Africa it will give a range and it can be quite a large range so you have a minimum that's very very different from the max so in that case how I'm going to say I guess I don't know this is something you develop the  
[04:53:52] framework but how Think about that in terms of story is that if you have this very wide range does that seem >> so in the theoretical framework I'm actually going to abstract from the range but in the setting they don't advertise a specific value they  
[04:54:07] advertise a range and throughout what I find is it's not that the large firms have a higher average salary because the range is more dispersed the range is actually more compact this distribution is shifted to the right and then that remains the case and the range does not  
[04:54:21] expand in response to the pay transparency mandate either and I think that has something to do with what I was saying earlier which is that firms worry that workers anchor on the max. So when you when you increase the dispersion that has some costs  
[04:54:35] uh if the firm chooses to hide it in the in the uh baseline before my experiment um it would be hidden from the workers.  
[04:54:42] Yes, >> small.  
[04:54:50] >> As I said, it's just a catch all proxy for good firms. Um, but I could use other, for example, if I use the your baseline salary before I introduce the experiment. I would see the same results just because firm size and wages are so  
[04:55:04] correlated. But the reason I don't do that is because then there are some new firms that enter and they just undermine my power.  
[04:55:17] Yeah. So in the paper I I do this whole exercise of figuring out what about a firm explains these characteristics like disclosure and I just find that this is the biggest predictor and it's not anything about the job itself or or the  
[04:55:31] types of workers that they're targeting. Okay.  
[04:55:35] Um, so in this framework, I want to think about what drives workers to choose a particular type of a job. And I'm going to do that based on the expected utility that workers plausibly derive from that job. So workers are making a decision about whether or not to apply to a job. There's going to be  
[04:55:49] two types of workers, male and female, and they're going to get uh utility from the wage component of the job and from the amenities. The amenities, I'm going to think just about the ones that women care about because I'm trying to explain that gender gap in search. Um, and so  
[04:56:03] it's going to be just about flexibility. And then in terms of the preferences, I'm going to assume for now that men and women have the same preferences over wages, but different preferences over amenities. The reason I do that is it's obvious if men and women care  
[04:56:17] differently about wages that pay transparency would then have different impacts. Uh, but what I want to show to you is that even in a world where they care the same amount about wages, this interaction with amenities can lead them to respond differently to information  
[04:56:31] about wages. Um so in the world where salaries are visible this is what their expected utility would look like. In a world where salaries are hidden the key argument of the paper is that they now cannot guess the salaries and they face instead a lottery over the salary  
[04:56:45] distribution and they will average their utility over this lottery. Yes.  
[04:56:56] >> Oh sorry at some point the salary is revealed and that's how the match and that's what the firm cares about. Do you find that the women do you find effects on actual job matches that women are more likely to get the jobs?  
[04:57:08] >> So, I'll show you most of my data and most of this paper is going to be about the job search because I want to think about whether workers make different choices just when I change the information. I'm not going to have as great data on hiring, but through a back of envelope exercise, I find that  
[04:57:21] women's hiring likelihood at these large firms that reveal salaries is 18% higher and men's actually goes down. And this effect is driven by competition. I'll show you. We'll get to that when when I show you the results.  
[04:57:36] >> Like I don't quite understand the risk of like you're going to get like you're going to find out the way before you take jobs, but why? Like >> so all of this is being modeled at a point in time where you're deciding whether or not to apply to the job. So  
[04:57:50] there's some behind this there's some cost of application. You're not applying to all jobs. And what my paper is saying is that there is quite a lot of you cannot hire a woman who hasn't even applied and there's just quite a lot of high wage jobs that women are not applying to. And this these decisions  
[04:58:04] that are made at the search stage end up mattering. And so that 18% increase is also actually coming not because firms change their behavior but rather your hiring likelihood was zero when you were not applying and it increases to about  
[04:58:17] 18% when you start applying and you're a bigger share of the applicant pool.  
[04:58:25] Yes, I could. Yeah. Um, okay. So, in this world where wages are not revealed, um, before I impose any additional structure on this very simple expected utility framework, there are two things. One Ben hinted at which is risk aversion. If the theta is  
[04:58:40] between zero and one, then risk aversion by Jensen's inequality implies that the utility of an uncertain wage is lower than u the utility of getting the average wage with certainty. Or to say that differently, uh when wages are  
[04:58:55] hidden, they enter the utility function with a risk discount if you're just a worker who doesn't like uncertainty. So that's something you can keep in your mind when you're thinking about well don't workers know that large firm salaries are higher? They may know the average. They may know the distribution.  
[04:59:08] But if the risk discount is high enough, that may not sufficiently compensate you for that higher average. The other implication of this very simple framework is that when wages are hidden, all the jobs that hide wages look the  
[04:59:22] same on the wage component. For workers to then rank those jobs uh within themselves, they have to do it on the amenity dimension. And that's how a worker with a given amenity preference can seem more amenity loving because now  
[04:59:35] that's the only uh dimension of the job on which they can discern between them.  
[04:59:39] And that's why hiding wages amplifies the role of amenities in the search decisions. That's general. But now let me impose a specific structure that I showed to you in the previous slide which is let me assume there are two types of jobs. Small and large firm  
[04:59:54] jobs. The large firms pay more, offer less flexibility and for simplification, let's say they don't reveal salaries, the small firms do. Yes.  
[05:00:05] Uh so I imagine that there's a joint distribution of amenities and wages by firm that people have an expectation over. Right. So I and I and I try to infer just by looking at the firm name or firm size or what other charact and  
[05:00:18] and then like these statements are they not conditional on whether the joint distribution like the crossorrelation is positive or negative. So yes, exactly.  
[05:00:28] This is the core assumption that I'm going to need in the model, which is that workers are not looking at the amenity dimension and figuring out what the wage is. If they knew that correlation, I guess it's also not sufficient to know the correlation. But when there is dispersion, like some  
[05:00:43] large firms pay more, some large firms pay less. and like knowing which realization you're up against matters.  
[05:00:48] But even the correlation I I'm not assuming that they know that because it's quite hard for them to infer the correlation based on cross-sectional job postings. On the one hand, um jobs that  
[05:01:01] have higher salaries will be able to afford more amenities and so there is that positive correlation across jobs, but then within the job obviously the wage is lower than it would be if the uh amenity was not associated with that  
[05:01:13] job. In most uh cross-sectional raw data sets um that correlation is going to be positive. So it's not and if workers observe a positive correlation they may look at large firms offering less flexibility and actually infer the wage is also low. So it's not really quite  
[05:01:28] clear how workers think about that >> they don't infer the conditional way the correct way conditional on a verb hiding  
[05:01:42] they they at the expected wage full distribution. Is that the idea?  
[05:01:47] >> I'm assuming that they do not read the non-wage characteristics and figure out what the right wage is. That's going to be what's going to drive most of my results. And I I validate that empirically by showing workers job ads from big and small firms within their  
[05:02:01] occupation and finding that workers are not able to read the text of the ad and figure out the salary even when I financially incentivize them to do so.  
[05:02:07] So that really is the core mechanism of the paper.  
[05:02:10] I mean just a comment is on in the theory section I think it would be easier to follow if you had the joint distribution and then told us what results are dependent on >> it sounds good. Yeah >> I think just so related to that um my  
[05:02:24] understanding then is from the sorting mechanism women are going to end up being sorted in equilibrium to lower pay higher amenity firms.  
[05:02:32] >> Yes thank and I'll just show that to you. Okay. So now I want to think about given the structure that I've imposed where large firms pay more, offer less flexibility, and are less likely to disclose their salaries, how workers choose which job to apply to. First  
[05:02:47] under perfect information, then under the status quo, where wages are not necessarily revealed. And what I find is that um in the in a world where there's perfect information, the expected workers are going to apply to a large  
[05:03:01] firm job if the expected utility of the large firm job is higher than the expected utility of the small firm job.  
[05:03:06] That's what that expression over there is saying. I can rewrite that in lots of different ways. I'm going to rewrite that in terms of the amenity preference, which is gamma because that's the only thing that varies by gender in this framework. And the idea is that you're  
[05:03:18] going to apply to a large firm job if your taste for amenity is below a cutoff that I'm calling gamma t. In the numerator of that cutoff is the wage advantage you get from going to the large firm. And in the denominator is  
[05:03:32] the amenity you lose from going to the large firm. And that's all weighted by how much you care about wages, which is theta. Now in a world where wages are not revealed, everything looks the same except now you have this um expectation  
[05:03:45] instead of observing a specific value. And what that does is it gives you a different cutoff uh which I'm going to call gamma c because it matches the control condition of my experiment which is very similar looking to the original  
[05:03:57] cutoff except instead of theta now you have this term theta tilda which is the original taste for wages minus this discount that comes from uncertainty. It comes from the fact that you don't observe a specific wage and you have to  
[05:04:11] do this pooling and potential risk discounting over the wage distribution.  
[05:04:15] The implications of that I'll show you in a second are that um the the two the cutoff um in the perfect information world is higher and I'll get to that in a second why that matters. But to repeat  
[05:04:28] the amplification point the relative weight that workers place on amenities compared to wages is going to be smaller in a world where there is full information relative to a world where there is partial information because  
[05:04:41] this theta tilda is smaller than theta. That's just to say in marginal rate of substitution terms, even though the preferences have not changed, information changes the relative weight that workers place on amenities and amenities start to matter more.  
[05:04:56] To think about this in graphical terms, um on the x-axis, let me put the amenity taste and on the y- axis the expected utility of the job. This gray line is the large firm's expected utility in my control arm where they don't reveal salaries. The green line is the small  
[05:05:10] firm's expected utility. they always reveal salaries and the idea is that you're going to apply to the large firm if your amenity taste is below this cutoff gamma C and when I mandate pay transparency the large firm utility  
[05:05:23] jumps up by a constant factor because that uncertainty discount is removed and so now there's a whole bunch of workers who start applying to the large firm jobs uh uh because the cutoff has  
[05:05:36] shifted up so I the first thing that I should find is that when I reveal salaries is not just that there's a wholesale increase in applications to all types of jobs, but if it's true that both men and women did not have information about large firm wages and  
[05:05:50] like more wages and potentially were also risk averse that both men and women should start sending more applications to these higher paying large firms. And second, if it's also true that they have differences in amenity tastes, then the  
[05:06:03] extent of reallocation should be gendered. In particular, what I'm interested in is that if I look at the gender distribution of the amenity taste and it is the case that women's uh preferences for amenities dominate men's  
[05:06:17] in this switcher region, then the extent of reallocation that I see from women should be larger than the extent of reallocation that I see from men. So those are the two things that I'm going to test. Do workers reallocate search towards higher wage large firms in the  
[05:06:30] presence of information? And is that uh does do women's responses dominate men's >> like the creative strategy for the firm  
[05:06:42] behind like a good idea for the firm what >> um so there's going to be heterogeneity it's going to depend on the I I talk about this in appendic section of the paper I think this deserves its own paper but the key argument is that there  
[05:06:57] is heterogeneity in firms bargaining and screening technology those are the two costs that firms tell me of pay transparency. We get too many applicants and we and bargaining is harder. There's also heterogeneity in returns to attracting talent. Like for  
[05:07:10] for some firms, it's valuable to get a 100 applications and find you know the most talented person. For others all workers are closed substitutes and you may as well get a small number of job applications, pick someone at random and  
[05:07:23] move on. Um so it's not clear. It it depends on what are the returns to talent and then what are the costs of screening and bargaining for a given firm. I think the fact that some firms change their behavior but not all  
[05:07:37] suggests that there is heterogeneity um in firms in response to that. But that's something that I want to explore in a different paper. Um okay so let me talk about the design of this experiment. In the control condition of my experiment um which is also the status quo firms  
[05:07:51] have to report the minimum and the maximum salary of that job. You see this red asterisk which is very helpful for me because this is a mandatory field for the firms to fill out when they're uploading a job post. But then they can optionally select this box saying let's hide the salary from the workers and  
[05:08:06] that's the status code that I maintain in my control condition in the treatment group. I simply take away this box and I say that as part of a reform that we're testing your job has been randomly selected. You have to post the minimum maximum salary of this job in the ad in  
[05:08:19] order for it to go online. Uh yes. Um so on in response to Ben's question so we're suppressing things about screening cost which is fine but there's  
[05:08:33] also this application cost which may or may not vary by gender. So we we haven't fully understood what the application costs are. So we can think about whether that varies by gender whether that can explain things.  
[05:08:42] >> Yeah. So absolutely I think uh the fact that all workers are not applying to all jobs there's that but then the the extent to which that cost binds can definitely be different by gender. I'm not speaking to it directly because I don't observe so much of the  
[05:08:56] intra-household bargaining situation, but what I imagine that cost looks like is you go to your father or your husband and you say, "Here are two jobs. One is flexible, but the salary is low. The other is not flexible, but I don't know the salary." And they tell you that,  
[05:09:11] "Well, let's do the job that's flexible and you're not the main bread winner.  
[05:09:15] It's fine." That's the kind of bargaining story that I think would make women's cost of application to a job that does not uh explicitly state the monetary returns of that job higher. Um but yeah, it's silent on all of that because I don't observe these household  
[05:09:29] dynamics. Yes. >> Um so the firms subscribe to a bunch of products the platform provides for screening applications once they have come in. And so for that reason it is  
[05:09:44] useful for the firms that the platform understands the salary range like throw out applications that are way out of this range. The platform even does some pre-screening for the firms uh in a subset of these cases.  
[05:09:57] >> They were on the screen. >> I can I can have a look at that. Yeah.  
[05:10:09] I'm curious you describe it as if there's a binary decision to apply to a large but in principle you could also use as a contin wages I only you know to price amenities  
[05:10:34] that would have been useful I only randomize whether information about wages is available But I think there are ways of doing it structurally and I can I can think about that. Yes.  
[05:10:51] and they might very differently than men know that >> that might get around a little bit the problem that you're having and it might be interesting to know time period  
[05:11:04] perhaps compare the range that they show when they have to show you know relative to the range that they were showing whatever a month Yeah. So I can do that exercise. I can do that exercise. I haven't done that yet. But I can tell you that the  
[05:11:19] treatment uh like when the firms are forced to disclose the wages that they disclose are not very different from the control group which suggests that they're not changing the salaries because on average  
[05:11:30] >> basically nonis. >> Yes. So >> one test like there are people who are applying men and women who are applying anyway to these women making more money.  
[05:11:44] like sort >> of so to get at the to get at the like to are the is there potential discrimination um at the firm side yeah I don't observe essentially what happens during the bargaining process but what I can tell  
[05:11:58] you is that women do not perceive large firms to be more discriminatory than small firms because I ask them what do you think you would get from this particular firm when I showed them a job ad what do you think the average woman would get what do you think the average man would get they don't site discrimination as a barrier for not  
[05:12:12] wanting to apply to that job. Um, yes. >> You know, you have that question about what do you think about firms who aren't  
[05:12:25] supporting the salaries, you know, then they might have a different answer. So, I think I think question is a varants large small  
[05:12:42] different. >> Yeah, I don't think that the respondents are necessarily responding to the firm size of these job ads. I think it's just that when you reveal wages, they can sort directly on the wages. When you don't reveal wages, they're sorting on  
[05:12:56] the amenities. And it so happens that the amenities that women care about are uh under supplied by these large high wage firms.  
[05:13:07] You're saying well do you see the large you think that large firms discriminate?  
[05:13:11] It would be a different question if you asked do you think that firms that don't disclose discriminate more >> I also do that I also do that and in the and in the in the lab so I also do a lab experiment uh which is a more controlled setting where I vary whether or not firm  
[05:13:26] size is mentioned whether or not some statement about where equal opportunity love to promote diversity is mentioned and then where the wages are mentioned and work from home is mentioned again I find I replicate all of these patterns so it's not these other uh perceptions  
[05:13:39] about the firm and the way in which these cor they correlate with each other because I cross randomize these these characteristics. Um so we're going to get to the alternative mechanisms. So let me first show you the worker results. Um the first headline result is  
[05:13:53] that when I mandate pay transparency applications go up by 49%. That's all jobs but importantly the results are concentrated in large firm jobs. So large firm applications go up by 66% small firm applications go up by 23%.  
[05:14:07] That resonates with the first prediction of my model which is that when we mandate pay transparency if workers are sorting on the wages then I should expect applications to only go up predominantly for the high wage large firms. The second prediction was about  
[05:14:22] gendered responses and what this bar graph is saying is that taking all of the jobs in my universe um the increase in female applications is 14 percentage points stronger than the increase in male applications. And specifically when  
[05:14:35] I focus on the large firm jobs the increase in female applications is 95% they nearly double and the increase in male applications is 59%. So again men also learn something new but there is this uh increased response from women  
[05:14:49] and I want to now figure out for the rest of the talk where it's coming from >> your prediction the model because you're showing us proportionally >> that's that was also proportional  
[05:15:03] because I was showing you those CDFs um okay so I've shown you that applications increase disproportionately for large firms and there's a disproportionate response from women now I want to convince you that it's coming  
[05:15:15] from the specific four mechanisms that I showed you at the start of the talk and not other factors like a change in perceived discrimination and so on. So first let me show to you that even during my experiment these four facts hold. So this first graph is going to  
[05:15:30] take the distribution and show you in the control group that large firms pay more. So the large firm minimum salary is 32% higher than the small firm minimum salary. The maximum is 21% higher. and to questions about the range. The range is more compact. Then  
[05:15:44] what happens when I mandate treatment? There were all these questions about whether that would change and what I find is that even after I mandate pay transparency, it remains the case that large firms pay more in the minimum and maximum and their range is more compact.  
[05:15:58] So that that fact is still there. Let me also show you that unless I force these firms to reveal salaries, it remains the case that during my experiment in the control group, large firms are going to be 19 percentage points less likely to  
[05:16:10] disclose salaries than small firms. So that information from the firm side has not changed. Now on the worker side, what do workers know about wages? So this is where I go to my survey. I sit down with the workers. that financially incentivized them to guess the salaries  
[05:16:25] of small firm and large firm job ads from their same occupation within 10% of the true minimum and maximum. And the key thing here is that workers don't really know salaries and that's true for both men and women. So the first uh red  
[05:16:40] bar is showing you that only 9% of women get the minimum salary right within 10% and only 7% of men do. And then that's the same like this lack of information prevails whether I'm talking about a  
[05:16:54] small or a large firm, a minimum or a maximum salary. So the idea is that both men and women are unable to read the text of the job ad and form a precise guess of the salary of that job.  
[05:17:06] And then if I shift to my experiment data, the idea here is to show to you that what changes is now they're able to respond to the wage component whereas previously they were not able to. So on the x- axis I have the wage on the y-axis I have applications and what you  
[05:17:20] can see is that the elasticity in the control group is flat for both men and women and then what changes is that now that they see the wage in the treatment group um there is this upward sloping uh elasticity for both men and women. So they start engaging in directed search now. Yes.  
[05:17:35] >> I mean just to clarify how much of this is the actual platforms algorithm like if you're searching for a job you put in the filter for wages.  
[05:17:43] How is that? >> Um, so if you put a filter for wages, then yes, you wouldn't see the jobs that don't reveal the wages. Um, I think the key thing is that there is this  
[05:17:56] differential at baseline, there was this differential uh search by gender and that isn't there once um I forced firms to reveal. So the idea is that it it  
[05:18:10] could be that only women were using that filter. I guess that would explain that fact. Um, and if so, I guess that's that's fine. Now, you start to see all of the jobs that have wages. But the key thing is you don't start applying to all of the jobs that have wages. There is  
[05:18:24] this gradient where you apply more to the high wage jobs. So, it's not just coming from you turning on a filter or turning off a filter, but within the jobs that reveal salaries, you discriminate on wages.  
[05:18:45] So yeah, so what this is showing you is that both men and women uh respond similarly when they see the wage information and in equilibrium that then means that previously women were not applying as much to the large firms and now the number of applications that  
[05:18:59] these firms are getting from women has increased.  
[05:19:02] Um, okay. It means that absent salary information, you're not able to direct more of your applications to the higher wage jobs.  
[05:19:13] Both men and women are not able to guess the salaries and just send more applications to job like >> I so I'm holding a constant here. I'm  
[05:19:29] holding everything else about the job constant. I'm trying to see whether or not you were able to guess the salaries of the job conditional everything else.  
[05:19:36] >> So fancy you were able to incorrectly guess or the salary are those people updating more and applying more after the treatment.  
[05:19:43] >> They are applying more. Um I don't think they're necessarily they are I think what I find is that they see the wage and then they react to the wage if it is higher. I don't want to make the claim that they now learn  
[05:19:57] something about the market correlations because I don't find that over time they get better at guessing the salaries of the control jobs. Uh there's still a lot of dispersion. They only seem to be able to respond to a wage once it's  
[05:20:10] advertised. Yes. >> Yes. If I here I'm holding amenities and everything else constant because I  
[05:20:23] wanted to make the point that women and men both are not able to direct search on wages when wages are hidden and when they become revealed differentity  
[05:20:41] because I value that I value that amenity at $2.  
[05:20:48] I think that's still possible. >> One thing I can do is I can make an index of amenities and I can put that on the x-axis instead of wages and then show that absent wage information. How  
[05:21:02] much do men and women respond to amenities? And then does that change when wages become disclosed? But that's not quite this graph.  
[05:21:10] Okay. So what I've shown you so far to summarize is that large firms hide wages and uh workers are bad at guessing the salaries. When the wages are hidden, they're not able to direct their search  
[05:21:23] in response to wages. And when the wages become visible, both men and women are directing search in response to higher wages. Similarly, so then where is the result that large firms then get more  
[05:21:37] female applications than before coming from? And my argument here is that disclosing wages then has this interaction with amenities. So first let me show to you that large firms underprovide certain amenities. Um I'm  
[05:21:52] going to focus on work from home because that ends up being the most important amenity that women respond to. And this graph is saying that in the control group small firms are twice as likely as large firms to allow you to work from home. And that remains the case in the  
[05:22:04] treatment group. And now let me turn to workers and show to you that women modestly care more about remote work but it becomes really amplified in the field. So I do a lab experiment where I show women and men two types of jobs  
[05:22:18] where I'm going to keep some information constant. Uh I guess I can't do that.  
[05:22:22] I'm going to keep some information constant and then vary whether first I mention this is a large firm with a 100 highly driven employees. whether I mention the salary range, whether I mention that this is a remote job, and whether I then have some statement about we value diversity, we're equal  
[05:22:35] opportunity, and so on. Crossrandomizing these all of these characteristics should get at the concerns that were raised around this idea that maybe revealing wages makes you update on other non-wage amenities as well. And I  
[05:22:48] don't find that that happens. The one thing that I find that is quite robust is that women care about remote work. So they are six percentage points more likely to apply to a job that's remote.  
[05:22:59] Men don't seem to care about it at all. The six percentage point finding, it's there, but it's ultimately quite modest.  
[05:23:07] And that's the argument of the paper that while the true preference in a very controlled setting is quite modest, when we go out in the world where there are information frictions, this preference becomes amplified. So now let me show to you what happens out in the world where  
[05:23:21] there are information frictions. Focusing on my control group where salaries are hidden, I find that women are 157% more likely to apply to a remote job than an on-site job. Whereas men are only about 65% more likely to  
[05:23:36] apply. So there is this very large difference in application behavior even though in the lab their preference for remote work didn't seem that large. In comparison, characteristics about the firm such as whether the firm is large  
[05:23:48] or not does not matter for women and it does on the margin seem to matter for men even when the salaries are not shown which I interpret as men are also directing search on some non-wage amenities but the non-wage amenities that men care about happen to be  
[05:24:03] correlated positively with firm size and stuff whereas the non-wage amenities that women care about are not supplied by these firms. This equilibrium changes when I mandate pay transparency. Now applications to remote jobs from women  
[05:24:16] go down by 73.5 percentage points. Applications to large firms in parallel go up by 43.5 percentage points. And in comparison, the changes for men's behavior are not that large or that significant. So the men perhaps who like  
[05:24:30] remote work just genuinely like remote work and we're not underpricing the cost of that.  
[05:24:37] Okay, we've talked about a bunch of alternative mechanisms. I'm mostly going to skip them now because I promised you some information about the firms. So, let me get to that. Uh, at baseline, I interviewed about 300 firms and two of  
[05:24:50] 200 of them were hiding salaries. First, I asked them, "What do you think would happen if you revealed your salaries instead of hiding them?" And 63% of the firms said, "We think it's going to increase applications." That's consistent with those firms think our  
[05:25:04] salaries are quite good. and when we advertise them, we're going to get more applicants. But only 28% of them think that it's also going to improve the quality of applicants. So they expect quantity to go up and they worry that  
[05:25:17] quality will not keep up to compensate us for the screening costs. This matters in a world where screening itself is costly.  
[05:25:26] The other thing that they report is about 71% of the firms say that the person we end up hiring may be unworthy of the maximum salary that we advertise.  
[05:25:34] So it undermines our ability in the bargaining process to tailor the salary to the productivity of that worker.  
[05:25:41] Despite these costs that these firms uh state, I find that after the experiment, firms become more likely to start voluntarily disclosing the salaries and that this behavior comes predominantly from the high wage large firms that were  
[05:25:54] initially hiding it. So uh in general, if you were never exposed to the experiment, you have a 63% likelihood of hiding the salary after the experiment.  
[05:26:03] that goes down by 11 percentage points and this is for all firms but this result is really coming from the large firms if they were never exposed to the experiment. Uh after the intervention they have an 81% likelihood of hiding salaries and that goes down by 25  
[05:26:18] percentage points if I had treated all of their jobs during the experiment. So to sum up the conversation with Ben earlier, it's clear that there are some firms that had overestimated either these costs or underestimated the gains in quality, but crucially not all firms,  
[05:26:33] which implies that there is some heterogeneity within the firms themselves. To speak to whether or not these costs are reasonable, uh whether or not firms have rational expectations.  
[05:26:42] Are these costs really there? I look at the CVs of the workers who apply to the large firms and compare their attributes to the job description itself and what I find is evidence of modest screening  
[05:26:56] costs rising in response to pay transparency. So first if I look at whether the average applicant meets the skills requirements of the job I find that that likelihood goes down by about 3 percentage points. If I look at  
[05:27:10] whether they meet the education requirements of the job, I find that that also goes down by about a percentage point. And then on all of the other attributes in the CV that I can compare to the job, I find similar magnitudes. So the idea is that at least on the observables, there are some  
[05:27:25] modest costs. The applicants who um come in as a result of pay transparency are on average a little bit worse, but not these costs are not massive. It's possible that on unobservable something  
[05:27:38] else is going on. So what I think of is firms are perhaps justified in being worried about this, but it's possible that they had overstated the extent of the concerns. Of course, you're not going to hire the average person. You're  
[05:27:53] putting this information out there to attract talent from the market. So the next thing I want to think about is if I construct an index using the rums of the workers of the best workers in the market, what is the likelihood that your job attracts those workers? And I think  
[05:28:08] what's really interesting here is that if I look at across these resume characteristics, the top desile of workers and think what is the probability that they apply to a control job. I find that in the control group,  
[05:28:23] one of these talented that at least one talented man applies to your job is nearly certain, but it's really that the most talented women were only 63% likely to apply absent salary information. And  
[05:28:36] when salary becomes revealed, um, their likelihood increases by 7 percentage points. So now, if you think about it from the firm's perspective, you just have to get one talented worker to do your job. And there and if you think men  
[05:28:49] and women are good substitutes, then you're already quite certain to find a good man even without revealing salaries. It's really on the female dimension that you're going to gain in quality if you reveal salaries. That's plausibly either something that firms  
[05:29:04] don't know or don't value. And that's why I think that they may have been reveal hiding salaries and then continue to do so exposed. There is this cost on the low abil ability dimension as well  
[05:29:17] where if I look at the bottom desile there is a small increase that some lowability women now start applying but the the key is that the increase in the high ability applications is larger than the increase in the low ability applications. is  
[05:29:32] >> on the idea that a worker is like good enough like if it was all kind of linear or whatever worker quality then I would care about getting more workers because I got more workers up but it's kind of is this really predicated the fact that  
[05:29:46] quality is >> that's why I think that it there's going to be heterogeneity by firms there are jobs where there are high returns for attracting talent like thinking about the economics job market you're going to filter through lots of applicants because finding someone good is really valuable but there are other jobs where  
[05:29:59] everyone is a close substitute and so it really doesn't matter. Okay, I'm down to the last 10 seconds. So, what I'm going to do is skip to the conclusion. Um, what I hope to have shown you today is that men and women care equally about  
[05:30:12] wages when the wages are visible. But when wages are not visible, they direct their search on non-wage characteristics. The characteristics that women care about are the ones that are overs supplied by the low wage firms and under supplied by the high wage  
[05:30:25] firms. So absent pay transparency, women end up applying more to these low-wage firms. When wages become disclosed, that inflated appeal of the amenities goes away and directed search on wages allows them to target the highwage jobs. This  
[05:30:39] suggests that policies that reveal pay rather than providing female friendly amenities can themselves go a long way at a low cost. Thank you.  
[05:31:19] I think so. All right. Thanks. Thanks a lot. Thank