Written By: Ciara Carl

April 17, 2024

This is a bittersweet moment as I write my last BESTEST blog post before graduation. With that being said, I feel that it is only right to write about my master's thesis, “Intersectionality and Employment Barriers: Analyzing Age and Gender Bias in Job Advertisements,” as I have been working on this throughout the entire program, with my blood, sweat, and tears poured into it. Well, maybe not the blood part, but you get the picture. I am going to walk you through it all - from choosing the topic, collecting data, refining methods, data analysis, main findings, and real-world implications of the study. 

Part of the two year M.S. in Psychological Science program is conducting, writing up, and successfully defending a master’s thesis under the guidance of a faculty advisor. To begin, I did not know the exact topic of my thesis until the end of my first semester. It is not until the second semester (Spring 2023) of our first year that you enroll in a course called Advanced Research Design (we call this course “thesis”). Dr. Walker had me reading research articles throughout my first semester on topics I was interested in before we began to brainstorm ideas/research questions. Please see my past blog on how I ethically used artificial intelligence (AI) to assist me with developing research questions. My research interests entail social psychology and topics related to implicit biases, stereotypes, and intersectionality, which is very broad and I know I needed to narrow this down a bit. Dr. Walker sent some articles she thought I might be interested in, and that is where my search began. 

Narrowing Down My Thesis 

Guiding Theories of the Study

1) Intersectionality: how two or more Social categorizations (e.g., race, age, gender) intersect in their impact on individuals or groups, creating overlapping systems of disadvantage. 

2) Social Role Theory: how the division of labor and societal expectations shape the development of individuals within different social groups.

• Used to explain why individuals pursue careers that align with their traditional gender roles (e.g., women as a nannies; men as firefighters)

3) Similarity-Attraction Hypothesis: posits that individuals are naturally inclined towards those who are similar to themselves

• Poses risk of affinity bias (i.e., similarity bias), potentially excluding those who are dissimilar to decision-makers. 

• Employers may unconsciously favor candidates who resemble them (based on demographics, personality traits, etc.), which may lead to affinity bias when creating a profile of an ideal candidate, leading to a less diverse applicant pool.

Now that I have given you some background information, let’s get back to how I developed the study. Being that a gap has now been identified in the current literature on this topic, I developed research questions to address these gaps (see poster below for my research questions). In our thesis class, we each began by drafting a research proposal to apply for UTC’s SEARCH Award (which stands for Scholarship, Engagement, the Arts, Research, Creativity and Humanities) to receive funding for our thesis project (e.g., money for equipment, to attend conferences to present findings, etc). I enjoyed how the class was set up and we did the SEARCH proposal first, as this served as a basis for our thesis proposal draft due at the end of the semester. Essentially, this class allowed for us to develop our literature review and methods for our actual proposal.  At the end of the semester, we gave a “mock” presentation to prepare us for the real proposal to follow.

Research Obstacles, Refining Methods, and Data Collection

Initially I wanted to conduct a two-part study, where part one focused on the attraction phase and part two would be a follow-up correspondence test to assess intersectionality (i.e., age and gender) in the selection phase. A little before my mock proposal, UTC IRB (institutional review board) came into our class to discuss ethical conduct in research regarding human subjects and also to discuss the different forms needed to be filled out based upon each of my cohort’s research. When briefly discussing our topics with IRB, they seemed hesitant about part two (i.e., correspondence test) of my study proposal given that employers would not be giving their informed consent. Please see my previous post, as I further discuss ethical concerns about correspondence tests in more detail. After talking with Dr. Walker, we decided we were going to get rid of part two of the study for my actual proposal. However, for the class “mock” proposal we still included part two, as I had spent a lot of time, conducted a lot of research, and had already written up my mock proposal. This ended up actually working in my favor, as I did not need to obtain IRB approval because I did not have human participants and my sample of job ads was online, publicly available data. Also, there has already been a decent amount of intersectional research (i.e., age and gender) conducted on the selection phase in hiring. Ultimately, our focus was the attraction phase, therefore this was not the end all be all. 

I proposed my thesis a month after the Spring semester ended, in June of 2023. After that, my research assistant, Will Higdon, began collecting 800 job advertisements (see below for the total sample of job ads) from LinkedIn and Indeed.com, storing them in a job ad database in Excel (included the job ad number, date collected, link to the posted job, location, employer name, number of employees, job ad text, and title of the job). While Will will was collecting the job ads, I was working on building a codebook/repository of stereotypes for age (young, middle, and older adults) and gender (man, woman), and gendered age (which was exploratory) based upon past research and literature. I just began putting all stereotypes I was finding for each category with the corresponding demographic (e.g., all older adult stereotypes under the older adult tab). 

Throughout this process, Dr. Walker and I realized we were going to need a different approach, as there are positive and negative stereotypes associated with each (going to confuse who was the ideal candidate if we did not differentiate). Therefore, we decided to split up within each excel sheet into “ideal” and “not ideal” based off the positive vs negative stereotypes. For example, one excel sheet was labeled “young adult” this was broken up into “younger adult ideal” and “younger adult not ideal.” 

My main message for explaining this, is that sometimes when conducting research you are going to have to make adjustments throughout and that it is a learning process in itself!

Data Analysis

By the end of the summer, Will finished data collection, and I began to analyze each job ad qualitatively through a content analysis. Initially, I was analyzing and coding the job ads for anytime I saw previously used language/phrases as well as adding new language/phrasing within the ads that was essentially saying the same thing, just in a different way (i.e., nuanced language). After coding the first 200 job ads (software engineer and business analyst), Dr. Walker and I reassessed this approach given the time it was taking to code even one job ad, as my codebook has a list of hundreds of stereotypes for each category (lol I am not kidding). After reevaluating, we decided to only pull out nuanced usage of gender and age biased language, as the second part was going to consist of a quantitative text-based analysis, using the software LIWC (Linguistic Inquiry and Word Count). Because I had already identified past language stereotypes for each category of gender and age, we were going to be able to implement this into LIWC, which would give us a frequency count instead of doing this manually. 

That is right, we found out that part one of the two-part study ended up having two parts within itself. If I would have done the initial two parts in the “mock” proposal study (with the correspondence test), it would have really been three parts….. I dodged a bullet 😅

For the majority of my Fall semester of 2023 (entering my second year as a graduate student), I spent analyzing the 800 job ads and pulling out the nuanced use of words. At midpoint, we also decided to remove middle aged adults from the age category as there was too much overlap between the younger and older adults, which made it challenging to differentiate which words/phrases belonged to the corresponding age “ideal” category. For instance, the word “experienced” has been found to be a stereotype of middle-aged workers (Finkelstein et al., 1995), causing difficulty determining its classification between older and younger adults. Other examples included “mature,” “responsible,” and “competent,” which also overlap with older adult ideal language. Also, there were middle-aged adult stereotypes and meta-stereotyped phrases such as “fresh, new ideas” and “innovative/creative,” which coincided with “young adult ideal” (Diehl et al., 2023).

After this step in the research process (exploratory content analysis), I received the licensing for the software LIWC through my SEARCH Award funding (as I was a recipient). Then my research assistant and I began to learn how to use LIWC through the user manual, watching several Youtube videos, and practicing with premade dictionaries and data sets available through LIWC. Dr. Walker looked over the words/phrases within my codebook prior to running the analyses in LIWC. Ultimately, we decided to remove explicit (i.e., conscious) negative stereotypes (e.g., worse at hearing, irresponsible, unadaptable, technophobic), as when looking through each job advertisement through the content analysis, negative stereotypes such as these were not found within the job ads. Stereotype congruent language (i.e., situations where information or behavior aligns with pre-existing stereotypes associated with a specific social category) was used to build ideal profiles. For example, researchers have found that older adults are stereotypically viewed as "unadaptable" (Koçak et al., 2022). On the other hand, younger adults have been found to be “adaptable (Centre for Ageing Better, 2021). So, terms that were stereotype incongruent towards one group (e.g., older adults being unadaptable) but stereotype congruent with the other group (e.g., younger adults viewed as adaptable) were deleted to only be associated with the positive “ideal” candidate demographic to conduct analyses (younger adult ideal). 

LIWC has a core component that lets you create your own dictionary categories. Then you can upload a file, and it will give you a percentage of the use of the text in the file that used words/language from the specific dictionary. For my study, each job had received scores (0-100) to indicate the level of both age-biased and gender-biased language, similar to the methods used by (Gaucher et al., 2011). The dependent variable (gendered age language) was calculated by multiplying the gendered language score by the age language score (man ideal and young ideal; woman ideal and older ideal). If the score was closer to 100, then this indicated an employer’s ideal candidate based on both age and gender. For example, a score of 38.3% on man and younger ideal wording, indicated that 38.3% of the total words in that advertisement were from the list of words within the dictionaries. After quantifying the words/phrases, I was able to conduct factorial ANOVAs to see how industry (white vs blue- collar) and job gender dominance (male vs female) intersect to signal an ideal candidate based on age and gender (e.g., young man ideal, older woman ideal). 

Key Findings

Bona Fide Occupational Qualifications

Feedback given on my proposal from one of my committee members who is an Industrial-Organizational Psychologist, Dr. Kristen Black, was to account for bona fide occupational qualifications (BFOQs). BFOQs refer to a characteristic or trait deemed essential for eligibility in a specific position, which may be directly related to job performance. BFOQs permit employers to select individuals based on characteristics that might otherwise be considered discriminatory, such as age, gender, nationality, and religion, if these traits are essential for specific requirements for the job. Essentially, if a job has a BFOQ, the exclusion of certain characteristics can be justified if it is demonstrated to be "reasonably necessary" to operate in a role. I am currently in the process of preparing my thesis document for publication, where I have went back and accounted for some of the BFOQs (e.g., the term “analytical” has been found by past researchers to be masculine skewed; however, this skill is necessary for the roles of software engineer and business analyst. Therefore, for the LIWC analyses for these occupations, this term was removed).

Real-World Applications of the Study

• • Job ads need to be framed in terms of behaviors and qualifications instead of fixed personality traits. Here are some examples:

Use “About the Role,” “Job Summary,” “Responsibilities,” or “Skills” instead of “About You,” “Who You Are,” or “Personal Attributes.”

• • Educate human resource professionals and employers to use age and gender inclusive language. Example:

“Experienced and new professionals’ welcome.”

• • Companies should invest in continuous training to assess job ads for potential unconscious biases.

• • Employers need to adopt an open mindset when searching for candidates to attract a diverse candidate pool with a range of background, knowledge, and experience.

Special Thank You to Dr. Walker