Sexual harassment in academe is underreported, especially by students in the life and physical sciences

What factors predict the underreporting of sexual harassment in academe? We used logistic regression and sentiment analysis to examine 2,343 reports of sexual harassment involving members of university communities. Results indicate students were 1.6 times likely to not report their experiences when compared to faculty. Respondents in the life and physical sciences were 1.7 times more likely to not report their experiences when compared to respondents in other disciplines. Men represented 90% of the reported perpetrators of sexual harassment. Analysis of respondents’ written accounts show variation of overall sentiment based on discipline, student type, and the type of institution attended, particularly with regard to mental health. Our results suggest that institutional and departmental barriers driven by power asymmetries play a large role in the underreporting sexual harassment among students—especially those in STEM disciplines.


Introduction
Sexual harassment in the workplace is pervasive and disproportionally experienced by women. The majority of sexual harassment reports made to the Equal Employment Opportunities Commission (EEOC), for example, have been initiated by women [1] and 60% to 75% of women report having experienced "unwanted sexual attention or sexual coercion" in the workplace [2]. Recent studies have shown similar patterns for college students-50% of female students experienced some kind of sexual harassment during their college years, and women of color have been shown to experience especially high rates of harassment [3][4]. In the academy, the majority of reported sexual harassment involves unwelcome physical contact (e.g., groping, sexual assault, and domestic abuse-like behaviors) by repeat offenders [5]. Women in the academy are also more likely to be the targets of inappropriate sexual comments from supervisors during fieldwork [6], and are often targeted when they are trainees [7].
Those who experience sexual harassment in post-secondary settings suffer mental, psychological, physical, academic, and work-related consequences-including lower GPA for students and, in the case of faculty, potentially leaving academe altogether [8][9][10][11]. The negative effects of sexual harassment, moreover, can affect an entire scholarly community; recent work a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 of the writers' opinions on a topic of interest (e.g., [30][31][32][33]). Chen, Zhu, Kifer, and Lee [34], for example, used sentiment analysis to compare the opinions of Republican and Democratic senators with respect to a variety of topics, focusing on similarities and dissimilarities between the two groups' opinions. Other work has found differences between men and women when it comes to the emotional content of work-place email [35]. Sentiment analysis has also been used to investigate the mental health status reflected through written language (e.g., [36]), and has been used with geographic information system data to identify communities in need of mental health services in the wake of natural disasters [37].

Current study
We contribute to studies on sexual harassment (e.g., [19,16,[38][39][40][41]) by focusing on reporting behaviors in the academy, and the sentiment present when describing harassment. Specifically, we analyzed unique reports of sexual harassment that occurred in higher education settings across diverse disciplines and institution types. These reports were collected through a crowdsourced survey, where the respondents shared the accounts of perceived sexual harassment through their answers to open-ended questions. We focus on characteristics that predict the reporting behaviors in order to identify potential trends across academic disciplines, and our analysis answers the following research questions: RQ1: Which respondent, perpetrator, and/or institutional characteristics to predict the non-reporting behavior of sexual harassment? RQ2: What is the sentiment of respondents' answers to open-ended questions about sexual harassment experiences?

Survey instrument
This study was approved in accordance with the University of Southern California's IRB guidelines (UP-18-00179). The data used for our analysis was gathered through an anonymous survey written and disseminated by Dr. Karen Kelsky-a consultant who focuses on helping individuals navigate the Ph.D. job market and advocate for marginalized groups within the academy. Her stated goal for the survey was ". . .for the academy as a whole to begin to grasp the true scope and scale of [sexual harassment] in academic settings," [42]. The survey was launched in November 2017 on Google Forms, and shared through social media posts. The survey did not have a formal close date; responses were made publicly available and were downloaded February 2018. The crowdsourced survey captured responses from many individuals (n = 2,343) from a variety of institutions all around the country, and beyond. The survey captures direct and second-hand experience of sexual harassment in the academy at various points in time in a person's academic career (e.g., undergraduate, graduate, faculty).
Demographic items. Respondents answered demographic questions that included their status during an incident of sexual harassment (e.g., graduate student); the status of the perpetrators (e.g., tenured professor); the gender of the perpetrator; the type of institution where the harassment occurred (e.g., "R1 institution"); and the respondent's field of study.
Reporting. Respondents answered one item that asked whether they reported the incident to their institution, and the institution's response if the incident was reported.
Open-ended questions. Respondents also answered four open-ended questions that focused on: 1) what happened, and when it happened-what we term the "harassment narrative"; 2) the impact the harassment had on their career; 3) the impact the harassment had on their mental health; and 4) the impact the harassment had on their choices or life trajectory.
Sample. The majority of respondents were from U.S. based institutions. The data captures instances of sexual harassment that may not have been reported to the institution where it occurred, as well as respondents' reflections on the impact the incident had on various dimensions of their lives. The data reflect 2,343 unique respondents and 9,372 harassment, mental health, career impact, and life trajectory narratives (combined).
Respondents. The survey did not formally ask respondents to indicate their sex or gender. Many of the respondents, however, used feminine pronouns (e.g., she/her) or otherwise described those who experienced sexual harassment as "women" or "female" in their description of sexual harassment. These responses were coded for the presence of such pronouns in the narrative statement. A second rater coded a sub-sample (n = 434) to calculate agreement; Cohen's κ was run to determine if there was agreement between the two raters' judgments. There was moderate agreement (73%), κ = .570, p < .001. Given thus, we infer that approximately 43% of respondents reported that women were the targets of the sexual harassment experience. The majority of respondents (70%) were students at various stages of their academic careers during the alleged harassment, (e.g., undergraduate or graduate); 20% were faculty, and 10% were staff (see Table 1 for further breakdown). Respondents came from different fields, including the humanities (47%), the social sciences (26%), the physical sciences (16.9%), and professional schools (7%) (see Table 2).
Reported perpetrators. The vast majority of reported perpetrators were men (92%), with 73% of them also identified as faculty. Faculty perpetrators had various supervisory roles (e.g., advisors), and 35% of perpetrators were identified as being tenured (see Table 3 for more detailed breakdown of perpetrator statistics). Location. Harassment took place across various institution types: 25% of respondents reported experiencing sexual harassment at R-1 research institutions, 40% at Ivy or "Elite" institutions, and 33% at other intuitions (see Table 4 in for a full list).

Data preparation Sentiment scores generation for open-ended responses.
Text data (RQ2) were processed in R using the Text Mining (TM) package [43][44][45]. This process enables quantitative analysis of text, and also removes common stop word (e.g., the, it, a) to facilitate analysis. We then used the "SentimentAnalysis" R package [46] to conduct sentiment analysis on each of open-ended responses of the data set; this process generates a sentiment score between -1 (negative) and 1 (positive) for each of the narratives using existing dictionaries. Sentiment variables were generated for each of the passages below. Harassment Narrative. Respondents to the survey were asked: "What Happened and When?" The mean word length of responses was 57 (SD = 70).
Impact on Career Narrative. Respondents to the survey were asked: "The Impact of the Harassment on Your Career." The mean word length of responses was 11 (SD = 15) Impact on Mental Health Narrative. Respondents to the survey were asked: "The Impact of the Harassment on Your Mental Health." The mean word length of responses was 11 (SD = 18) Impact on Career Trajectory Narrative. Respondents to the survey were asked: "The Impact of the Harassment on Your Life Choices/Trajectory." The mean word length of responses was 10 (SD = 17).
Manually coded variables. We manually coded discipline and institutional response variables since many of them (e.g., the respondent's status within the university) were written as free responses in the survey, instead of pre-determined categories that respondents could select from. These variables were manually coded using the process below.
Discipline. Disciplines reported by respondents were aggregated into 6 different categories based on the National Science Foundation's categories of earned doctorates [47] (see Table 5.) Institutional response. The survey question read: "Institutional Responses to the Harassment (If Any)." Responses were coded into six categories based on existing literature, e.g., retaliation, taking actions to redress the situation, and not taking any action (e.g., [18,48]). A second rater coded a random-sample of 10% of the responses to calculate agreement; Cohen's κ was run to determine if there was agreement between the two raters' judgments. There was moderate agreement (64%), κ = .512, p < .001. These coded categories were (see Table 6 for more examples): • None: Respondents did not report a formal response by the institution, e.g., "None," "Nothing has been done," and "No formal response." • Did not report: Respondents did not report the incident, e.g., "Never reported it," "I didn't say anything," and "I never told anyone." The two largest categories ("Elite Institution/Ivy League" and "Other R1") were maintained for analysis. All other categories were combined into "Other." Note: "redacted" category from original survey results; the research team did not redact any data. https://doi.org/10.1371/journal.pone.0230312.t004 • Action taken: Respondents reported an institutional response after the incident was reported, e.g., "Title IX investigation," "He was fired," "He was removed." Did not report No reporting system for faculty/grad student interactions at the time; Discouraged from reporting due to potential damage to my own career; Never reported it; I didn't say anything; I did not report it; I did not report because I feared direct retribution; I never told anyone about this; Warned that it would get ugly if I filed a complaint so I didn't; Not applicable as happened at conferences 3 Action taken Title IX investigation; He was fired from his tenure position; There is supposedly "a file" on the incident somewhere; He was disciplined; This was enough to force the perpetrator to "retire" with immediate effect; He was removed from my tenure committee and was told to have no contact with me; Forced resignation 4 Unclear, Not sure I heard stories about "talkings to" that he received but I don't know; Dean collected information anonymously from myself and other victims; My hope is that this curbed his behavior significantly but I can't know for sure; I think he was rebuked; No information whether action was taken; Unclear; Unknown; I think there was an investigation but I never heard of any results; They spoke with him beyond that none to my knowledge • Unclear, Not sure: Respondents' indicated an unclear institutional response, e.g., "Unclear," "Unknown," "I think there was an investigation but I never heard any results." • Retaliation (against the respondent):. Respondents reported retaliation against them for reporting the incident, e.g., "I was silenced and punished," "I was fired," "Retaliation." • Other: Responses that do not fall into the above categories, e.g., "There were conflicting responses from faculty," "bathroom renovation over a year after the fact." Respondent status. Respondent status at the time of the incident was coded into 16 categories that ranged from graduate student to tenured faculty. These categories were based on the open-ended responses provided on the survey by the respondents (see Table 7 for categories and examples). Perpetrator's status. The reported perpetrator's status was coded into 8 categories, based on the open-ended responses provided on the survey by the responders. Some perpetrators had multiple roles (ex: a professor AND a dean) or relationship with the responder (e.g., a professor AND an advisor, see Table 3.)

Data analysis
Logistic regression model specification (RQ1). We used respondent, perpetrator, and institutional characteristics to estimate the likelihood of subgroups' choice to not report sexual harassment. We specified a logistic regression model (Eq 1, below) with perpetrator characteristics (β 1-9 ), respondent characteristics (β 10-15 ), and institution type (β 16 ). Variables were simultaneously entered into a logistic regression model using Stata 15.
The primary coefficients of interest were β 10-15 , which represent disciplinary affiliations of respondents. Coefficients of secondary interest were β 1-9 , which represent perpetrator characteristics (e.g., tenure status). Institutional type was operationalized as a control variable (β 16 ), since the original survey categories do not follow traditional Carnegie Classifications for institutions of higher education [49]. We are, moreover, unaware of extant literature that speaks to a pattern of institutional differences regarding sexual harassment reporting and would thus warrant a different approach.
Sentiment analysis (RQ2). Our analysis used dictionary-based semantic annotations to assign a sentiment score ranging from -1 (negative) to 1 (positive) to each of the open-ended responses. This approach was the most appropriate, given that we do not know of a collection of text responses that we could use as a reference category to use more sophisticated methods [50]. Different sentiment analysis scores suggest different emotional states and intensity of language used. Descriptive heatmaps were generated by plotting sentiment scores along two dimensions: 1) institution type, and 2) demographic characteristics of respondents. Heatmaps in the results section were generated using R. (See Table 8 for summary statistics of sentiment scores.)

Institutional response
Institutional responses to respondents reporting sexual harassment included: some action taken (9%), no action taken (34%), retaliation (8%), and unclear if action taken (6.5%). Choosing to not report the incident at all was the most common response (36%). This decision varied across disciplines, with roughly 50% of respondents in the physical sciences and engineering choosing to not report the incident. In contrast, only 18% of staff chose to not report the incident (see Table 9).

Logistic regression
Results from logistic regression analysis indicate that students were 1.6 times more likely to not report their experiences when compared with faculty and staff respondents (p < .001). Students in the physical sciences were 1.7 times more likely to not report their experiences when compared to students in other disciplines (p < .05). If the perpetrator was identified as a faculty member, then respondents were 1.5 times more likely to not report the incident compared to respondents who identified the perpetrator as a graduate student, postdoc, or other non-faculty (p < .001). Institution type (e.g., Elite/Ivy, Other R1) was not predictive of respondents choosing to not report sexual harassment (see Table 10 for full logistic regression results).

Descriptive sentiment analysis
Descriptive analysis of the open-ended responses (RQ2) indicated suggests variation in the mental health narratives, especially for students attending elite/Ivy League institutions as compared to other respondents (Fig 1C). A further breakdown analysis indicated that this pattern varied most among undergraduate and graduate students who attended elite/Ivy League institutions ( Fig 2C). Sentiment scores of the mental health narrative were more negative for students in the social sciences who attended elite/Ivy institutions when compared to students attending other institution types of institutions, regardless of discipline (Fig 3C).

Limitations
We note that the data that underpin our analysis were taken from a crowdsourced survey that was primarily distributed via social networks (e.g., Facebook), and as such likely suffers from selection bias. Specifically, it is likely that respondents to the survey have a tendency to use social networks regularly and to engage in social networks. Respondents were also likely familiar with None = no institutional response, unclear if reported; Did not report = respondents indicated incident was not reported; Action Taken = Reported, some action was taken by institution; Unclear = unclear if the incident was formally reported, unclear response to reporting; Retaliation = reported retaliation against respondent.

PLOS ONE
Sexual harassment in academe is underreported Dr. Karen Kelsky's service and/or somehow connected to one or more of her social network accounts. This suggests that, as a group, they may have a similar set of experiences [51]. This is born out in the data itself; nearly half of the respondents are from Humanities (Table 2) and other members of academia were underrepresented (e.g., administrative staff). Some respondent characteristics (e.g., race, sexual orientation, gender, age at the time of incident) and key dates (e.g., the year of the incident), moreover, were not captured by the survey. This is unfortunate, since such information would have enabled us to examine how reported experiences differed as a function of individual's demographic characteristics and/or policies that would have varied over time (e.g., the 1980's vs. today). Respondents' decision to report-and the reporting procedures available to them-have likely differed over time as a result of changes to EEO and Title IX policies and interpretations. In addition, retrospection bias (e.g., [52][53]), affects the reliability of the respondents' accounts. As such, we note that we do not make strong claims regarding the intensity of respondents accounts; our logistic regression analysis, for example, is limited to a subset of reporting behaviors (i.e., not reporting). Our covariates were chosen similarly so that we could mitigate against the influence of any retrospection bias. There are some challenges associated with the use of sentiment analysis as there are with all methods. One of the challenges pertains to the context of the text such that a particular word's meaning might vary depending on the context (e.g., a word that is positive in one context might be negative in a different context), or a lack of context in texts [54]. It is also difficult to detect subtle and implicit social cues such as humor and sarcasm, as well as inconsistencies in written texts-such as contradictory statements [54]. We emphasize that our sentiment analysis is descriptive, as we did not use scores in subsequent inferential statistical analysis.

Discussion
Our study supports to the notion that when sexual harassment occurs in the academy, it often does so via power asymmetries, i.e., scenarios which the perpetrator of sexual harassment is in a position of power relative to the person who is sexually harassed. Specifically, student

PLOS ONE
respondents were more likely to avoid reporting sexual harassment. Similarly, despite being underrepresented in our sample, respondents who identified as staff reported more retaliation (Table 9). Previous work has shown that power imbalances can elicit situations associated with high rates of harassment and underreporting [1,10,55]. Newins & White [16], for example, found that over a third of students (36%) were unsure whether they would tell a faculty member about sexual assault, and about a fifth (16% to 22% percent) were not willing to disclose at all.
The variability of underreporting behaviors across disciplines also suggests differences in disciplinary culture-the fact that only 50% of respondents in STEM fields chose to report their incident suggests that there are potential barriers to doing so, and may also suggest a process that makes reporting undesirable, a point underscored by the #MeTooSTEM movement and previous work [22,56]. Representation likely plays a role, as decades of work shows that women are underrepresented in STEM (e.g., [57][58][59][60]). Low representation of women, in addition to higher rates of experiencing sexual harassment, may thus create a difficult environment for women in STEM.
Sentiment analysis of written accounts helps to uncover additional information within a data set ([36-37]), and our descriptive analysis suggests institutional differences among the mental health narratives (Figs 1-3). These differences were most apparent when examining sentiment scores of students who attended social science programs within elite/Ivy League institutions (Fig 3). We infer from this analysis that one must also attend to institutional differences when investigating sexual harassment and its impact on the academic community. Such studies should attend to differences in implementing and reinforcing policies regarding sexual harassment [48,[61][62][63]. However, our study cannot definitely state that the apparent institutional differences in respondents' sentiments regarding the mental impact of sexual harassment were due to true institutional differences or were in fact an artifact of how data were collected. Regardless, it is imperative for each institution to provide pathways that can best support those impacted by sexual harassment and provide necessary services, such as counseling. Compulsory reporting practices, if in place, should also be examined so that the wellbeing of those who report is protected.
It is important to address sexual harassment in the academy; those who choose to sexually harass their students and/or colleagues are responsible for a host of negative outcomes among those they victimize. Chronic sexual harassment has been shown to be predictive of anxiety, depression, and substance abuse; negative health outcomes include hypertension and poor sleep [64][65]. Sexual harassment is also associated with psychological duress and lower academic satisfaction [66]. Men are not immune from sexual harassment in academia, but experience it at lower rates compared to women [4].
Recent work has highlighted the fact that women feel more supported and/or empowered to speak of when they experience sexual harassment [67]. Notably, one of the themes that emerged in the Keplinger et al. [67] indicates that the #MeToo movement has helped many feel comfortable with sharing sexual harassment experiences. Troublingly, however, while sexual harassment has decreased in the wake of the #MeToo movement, gender harassment (defined as negative views of women and gender hostility) has increased [67]. Universities should thus continue to pay close attention to gender harassment to ensure that it does not increase as sexual harassment decreases. Future work in this area might also examine differences between the reported sexual harassment experiences of women before and after the #MeToo movement. It is also crucial for future work to examine the reported sexual harassment experiences of populations may be particularly vulnerable to sexual harassment, such as LGBTQ community, people with disabilities, and people of color. Such work could reveal whether the institutions have made effective and tangible changes to the ways they handle sexual harassment. Taking such steps will help to ensure that the next generation of scientists are trained in an environment where safety and respect are the norm.