Tag Archives: things I will regret writing

SSC Survey Results: Sexual Harassment Levels By Field

[content note: sexual harassment]


Recent discussion of sexual harassment at work has focused on a few high-profile industries. But there has been relatively little credible research as to how rates really differ by occupation type.

There are many surveys of harassment rates in specific industries, but they can’t be credibly compared with one another. The percent of people who report sexual harassment varies wildly from survey to survey – thus studies finding that anywhere from 12 percent to 48 percent to 60 percent to 85 percent of women have been harassed at work. If a survey shows that 60% of female nurses get sexually harassed at work, does that mean nurses are victimized particularly often (because more than 12%) or are unusually safe (because less than 85%)? It doesn’t matter, because another study says only 19% of nurses get harassed.

Why do all these numbers differ so dramatically? The most important issue seems to be how you ask the question. “Have you ever been harassed?” gets numbers more like 12%; giving a long list of specific behaviors and asking “Have you ever experienced any of these?” gets numbers closer to 85%, depending on what the behaviors are. Surveys also differ on whether they ask all employees or just women, whether they include a time frame (eg “…in the past two years”), whether they specify that it had to be at work vs. work-related events, and whether they include witnessing someone else’s harassment. Taking these surveys entirely seriously would lead to the conclusion that Uber has the lowest sexual harassment rate of any company or industry in the world; I choose not to take them seriously.

This means we need investigations that use the same methodology across multiple fields. Whenever the media talks about this – see eg the Washington Post’s The Industries With The Worst Sexual Harassment Problem – they’re working off of the Equal Employment Opportunity Commission’s records. But these are totally unsuitable for the task – they just report raw number of claims per industry. The industries that rank lowest in EEOC’s data tend to be small industries with very few women – for example, taken seriously the WaPo’s graph shows that mining has the least problem with sexual harassment of any industry in the world. Is this thanks to their uniquely progressive culture – or because there are practically no female miners? I’m going to say the second one. The takeaway that most real researchers take from the EEOC claims is that the lowest-paying and most mundane occupations – retail, restaurant work, hotel work, etc – have much higher sexual harassment rates than the prestigious occupations people generally talk about. Eyeballing the data, this looks basically true. But trying to get anything more fine-grained than that out of EEOC is basically hopeless.

I only know of two surveys that have even attempted to compare different fields in a principled way, and neither really inspires confidence.

First is a survey by Cosmopolitan magazine, that asked 2235 women in different fields about sexual harassment in their industry. About 33% reported being harassed at work. Women in the retail and restaurant industries reported the highest rates of sexual harassment; despite media coverage’s focus on STEM and entertainment, both reported average to just-below-average harassment rates.

Second is the Project XX Survey, where the automotive industry decided to survey their workers using methodology previously used in Silicon Valley, which made their results at least somewhat comparable. The advertising and market research industries seem to have joined in later. They find the automotive industry has the most harassment, followed by Silicon Valley, advertising, and market research – but there aren’t really big differences and industry rankings frequently reverse across questions.

I interpret existing data in this area as being basically useless, but at least suggestive that the media focus on a few prestigious industries is mistargeted.


The Slate Star Codex survey is an online poll of readers of this blog about various aspects of their life and personality. This year 8,077 people participated. The survey asked a set of questions on sexual harassment, providing a unique dataset with which to investigate the question.

Due to the blog’s tech focus, readers skew well-off, male, and STEM-focused. This made it impossible to investigate the restaurant and retail industries that previous research suggests have the most problems. But there were still enough responses to get a wide cross-section of white-collar occupations and academic fields.

Respondents were asked four questions:

1. Have you ever been sexually harassed or assaulted at work?

2. Have you ever been sexually harassed or assaulted outside of work?

3. Have you ever made unwanted sexual advances to someone at work that you think they interpreted as sexual harassment or assault?

4. Have you ever made unwanted sexual advances to someone outside of work that you think they interpreted as sexual harassment or assault?

Respondents were given a “prefer not to answer” option in case the questions made them upset; in order to make people entirely comfortable using the option, they were asked to select it automatically if the last digit of the time was a “1”.

Data were analyzed about female victims, male victims, and male perpetrators. Although some women admitted to perpetration, the sample size across fields was too small to be useful. Industries were included if they had a sample size greater than thirty; because there were more male respondents and so all industries had a greater male sample size, some industries are included on the male graphs but not the female graphs.


Overall, 21% of women and 6% of men reported being sexually harassed at work. 1% of women and 2.2% of men reported committing sexual harassment at work.

Fields listed do not exactly match fields on the survey as some were combined in order to increase sample size, eg “physics” and “chemistry” into “hard sciences”. There’s a discussion involving pre-registered categories below.

Here’s all three measures combined into my best guess for the relative rates in each field:

Made by taking each of the three previous graphs, standardizing each field to percent of the worst field in that category, and then adding them up and dividing by three to get an average for each field

But this combines three different things. First, the actual rate of harassment. Second, a measure of how willing people are to report harassment. And third, a measure of how willing people are to consider some specific act to be harassing. This is a major problem. People from traditionalist cultures and subcultures may have a higher threshold for calling something harassing; if (for example), more traditional people go into Business, and more socially liberal people go into Art, that could skew the numbers.

In order to control for this and other factors, I asked subjects how often they are harassed outside of work. Of note, men, women, victims, perpetrators, everybody in every industry, reported much more sexual harassment outside of work than in it. Although there were enough female perpetrators to do statistics on this time, I left them out to keep it as similar to the in-work data as possible.

Here’s an overall summary, made through the same combination process as Figure 4.

Levels of at-work harassment and out-of-work harassment correlate at 0.81 across fields. This suggests that most of the differences observed here are differences related to the people in a field, rather than to any sort of field culture. Given how strong this effect is, it’s unclear whether it’s possible to adjust for it and get a measure of field culture. If we try (by dividing the at-work harassment levels by at-home harassment levels), we get this:

The most striking finding I find on all these graphs is that “nerdy” / STEM / traditionally-male jobs have the least harassment. The less nerdy / more verbal-personal-skills / traditionally-gender-balanced jobs have the most harassment.

Here’s Figure 4 again, but this time with nerdy-STEM-male occupations in red, verbal-personal-balanced occupations in blue, and edge cases in grey.

If categories like nerdy/STEM/traditionally-male seem unnatural, that’s fine – these are always a judgment call. But in December, I preregistered which clusters I would use, to prevent me from making them up to get the result I wanted. It’s not exactly the same as the clusters above – I ended out having to combine a few to get a big enough sample size to be meaningful. But using the original pre-registered clusters and the original pre-registered endpoint: in STEM/male-type fields, 12% of women (n = 212) reported being harassed at work. In non-STEM/balanced-type fields, 25% of women (n = 161) reported being harassed at work. p = 0.002. The preregistered analysis confirms the finding above.

I preregistered the hypotheses that STEM would have more female victimization (because there’s a higher ratio of heterosexual men to heterosexual women), but the same amount of male perpetration (because there is not an inherent difference in the sort of people who go into the field). Both preregistered hypotheses were wrong; female victimization and male perpetration were less.


Overall victimization rates were comparable to past surveys. These were on the low end of reported results, likely for several reasons. First, we strongly encouraged participants who were at all uncomfortable to say “prefer not to answer”, including asking 10% to say this outright regardless of experiences; this clearly depressed results and makes absolute numbers inaccurate. Second, participants were generally higher-income, and this group usually reports less harassment. Finally, we asked participants only about “sexual harassment” in general, and not about specific misbehaviors. Given these issues, the modest differences between these and past results (eg our 21% average female harassment rate compared to Cosmo’s 33%) are not surprising.

More surprising were the lower harassment rates in STEM and male-dominated fields, for two reasons. First, if most harassers are heterosexual, then a high percent male workforce in an industry should create a situation where a large number of harassers are focused on a small number of potential victims, raising the average harassment rate per victim. But this was not observed in our survey, nor in the Cosmo survey, nor (as far as can be eyeballed from the not-really-inter-comparable data) in the EEOC survey.

I don’t know of a simple explanation for the discrepancy. A friend suggested that it might be a customer effect. That is, lawyers / health-care-workers / etc often have to deal with sexist customers, whereas female programmers might not encounter anyone on the job except their coworkers. But this seems unlikely to make much difference; this survey finds that customers represent only 9% of sexual harassment incidents. Also, remember that the rates at which men admit to perpetrating sexual harassment match the rates at which women admit to experiencing it, and there’s no reason why the male-perpetration rates should have anything to do with customers. Also, social scientists don’t have customers any more than hard scientists do, but the social science harassment rate is higher. Perhaps the sort of people who go into STEM are socially oblivious in a way that prevents them from both noticing if they are sexually harassing someone else, and from noticing if they are sexually harassed. This would result in somewhat lower self-reported perpetration rates, and self-reported victimization rates that are even lower than what the low perpetration rates would predict. It would also match the high correlation between in-work and out-of-work harassment, which suggests that demographic factors are making most of the difference.

A second reason the low harassment rates in STEM were surprising is conventional wisdom claiming the opposite. In media coverage, STEM fields are portrayed as full of “techbros”, and uniquely unwelcoming to women. How do we square this with the data?

Some people who reviewed these data suggested that the media focuses more on a narrative where STEM workers are dismissive of women, rather than one where they harass them. I don’t find this accurate – a quick glance at any media source will show they focus on both narratives. But it seems possible that the dismissiveness narrative is correct, and the harassment narrative is a side effect of this. But this is somewhat contradicted by the auto industry survey mentioned above, which found tech was unexceptional either in harassment or in general dismissiveness. However, the other industries participating in the survey may themselves be unusually bad.

Another possibility is that the respondents to the survey may be skewed for some reason. For example, they may have been selected by this blog’s previous discussion of these issues, which have also argued that tech is not a uniquely bad industry.

But a final possibility is that the media coverage is inaccurate. The minimal amount of previous research in this area has all shown that the highest rates of sexual harassment are in the retail and restaurant industries. The media has ignored this. Before last year, the media almost never mentioned sexual harassment in Hollywood; now they hardly mention anything else. Given that already know media coverage here is generally inaccurate, maybe the data don’t need to be squared with it. Maybe the media is just wrong.

Every industry has enough sexual harassment to produce horror stories. If the media is disproportionately interested in the horror stories of relatively well-off people, and in stories that confirm their existing prejudices, we might hear many genuinely horrible stories about harassment in technology, but fewer about harassment in eg health care, philosophy, or law – even if objectively those fields are much worse. That might further encourage people in technology to come forward (consider what happened in Hollywood after the Weinstein revelations) and people in other fields to stay silent, contributing to a vicious cycle.

Compare all the anecdotes and popular lore about how immigrants are criminals. It’s totally false – immigrants have crime rates well below native-born citizens. But we only know that because there have been really good studies. If the studies hadn’t been done, and all we had to go on was the daily lurid stories about a Mexican guy knifing someone, who would believe it? In the absence of real studies, the media’s ability to spin a compelling narrative casting some people as monsters feeds on itself forever. We know that happens relatively often. Are we sure this time is different?

There’s only been one strong previous survey on this subject – the one in Cosmo I referenced earlier. They found the tech industry had an around-average to slightly-less-than-average rate of harassment. Anything else you’ve heard – any news story, any anecdote – is at the same level of evidence as the stories of immigrant crime.

It seems possible that this narrative grew up to explain why there are so few women in technology – maybe they’re all being harassed out by lecherous professors and bosses. But this explanation doesn’t hold water – the rate at which men vs. women choose tech courses in high school exactly parallels the rate at which they go into tech occupations. And gender balance in fields isn’t predicted by harassment levels in those fields anyway – otherwise you’d never see women working in restaurants, retail, or Hollywood.

I have been expecting results like these even before I did this survey, and this calls my data into question. Whenever someone with an unusual opinion produces an experiment that confirms their opinion, you should always be skeptical until it is verified by other sources.

So aside from appealing to the unimpeachable authority of Cosmopolitan as partially confirming my results, I urge anyone with the relevant skills to download the raw survey data themselves and see if they can replicate my conclusions. You can get them as .xlsx or .csv. Some people have complained of weird problems in the csv format and I recommend the xlsx if at all possible. I have removed the data of a few people who did not want their answers to be public, so you may not get exactly the same numbers I did, but they should be pretty close.

And more important, I appeal to anyone with an interest in this topic to do larger and more formal surveys comparing different fields. This may be the single topic where the extent of public interest is most disproportionate to the minimal amount of good research done. If people more prestigious than me or Cosmo were working on this, we would be in a much better position to know what to believe.