Manager Gender and Employee Preferences: Evidence from a Survey Experiment

We implement an online survey experiment to study individuals’ preferences for the gender of managers at work. Specifically, we want to understand: a) whether workers value male and female direct supervisors differently and if these preferences differ by worker gender; and b) whether these preferences are associated with gendered beliefs about managerial characteristics.
We recruit participants online via the Prolific platform and ask them to complete a Qualtrics survey containing a multi-stage experimental design.
After an initial set of background information, participants complete a discrete choice experiment (DCE), in which they choose between 2 alternative hypothetical jobs. Each job differs in terms of wage and a set of non-wage characteristics, with all attributes randomly selected. Overall, they make 10 choices.
The following stage elicits participants’ perceptions about a set of seven managerial characteristics. First, they rate each attribute in terms of how important they consider it in a direct supervisor. Next, for each characteristic, participants state whether it is most likely to be attributable to a female or male supervisor.
In a third stage, participants are given information about the gender composition in the highest earning quartile of a big company and are asked to state if they would be interested in finding out more about job opportunities at such company. The company can have a high share of either women or men in the top earnings quartile and participants are randomly assigned to it within their industry (or desired industry) of employment. If they state their interest, they are also shown on a separate screen a link to the company’s vacancies website, and we record whether they click on it.
Following the experimental stages, participants state their preferences for the gender of a direct supervisor and the gender composition of immediate colleagues at work. Finally, they answer additional demographic and current work-related questions.
This design allows us to estimate the willingness to pay, and hence individuals’ preferences, for the hierarchical structure of companies, and specifically the gender of the direct supervisor. In addition, we can understand whether such preferences are influenced by differences in perceptions about manager characteristics.

2026-07-03

2026-07-31

1. Willingness to pay for supervisor gender
2. Click on link to company website
3. Attribution of positive managerial characteristics to male supervisor

1. Willingness to pay for supervisor gender:
Using the jobs selected by participants in the discrete choice experiment and OLS estimation, we will estimate the willingness to pay (WTP) for the job attributes, expressed in % of wage terms. The main WTP of interest is the one for the gender of the direct supervisor.

2. Click on link to company website:
Participants are asked to state their interest in finding out more about job opportunities at the randomly assigned company in the third stage of the experiment (secondary outcome below). If they say yes, they are presented with a link to the assigned company’s website of job vacancies. Our outcome of interest is the binary variable which equals 1 if a participant actually clicked on the link to the company website, and 0 if participants did not click on the link or stated no interest in the first stage.

3. Attribution of positive managerial characteristics to male supervisor:
Binary variable indicating that the male supervisor was chosen as most likely represented by positive characteristics in the manager perceptions stage.

1. Stated interest in working for the company
2. Stated preference for the gender of the direct supervisor at work
3. Stated preference for the gender composition of immediate colleagues at work
4. Willingness to pay for the colleague gender composition

1. Stated interest in working for company:
Binary variable indicating if participant stated their interest in finding out more about job opportunities at the randomly assigned company in the third stage of the experiment, prior to being presented with the link to the company website.

2. Stated preference for the gender of the direct supervisor at work:
Set of binary variables indicating the preferred gender of supervisor or no preference regarding supervisor gender.

3. Stated preference for the gender composition of immediate colleagues at work:
Set of binary variables indicating the preferred gender composition of immediate colleagues (majority females, majority males, similar share of females and males) or no preference for the gender composition.

4. Willingness to pay for the colleague gender composition:
From the discrete choice experiment, we will also estimate the willingness to pay for the gender composition of the immediate colleagues at work.

Our experiment consists of a multi-stage only survey with experimental components and demographic and work background questions.

The demographic and work background information asked are split into two sets, one at the very beginning of the survey and another one after the experimental part, to avoid fatigue effects on answers to main stages.
At the beginning, participants are asked:
- age, gender (key moderator variable of the analyses) and ethnicity
- education level
- employment status
- industry they work in / want to work in and respective occupation
After all the experimental stages, participants are asked:
- income level
- political views
- caring responsibilities and number of children (if any)
- current work seniority (if employed)
- current gender composition of colleagues (if employed)
- gender of current supervisor/manager (if employed)
- any managing experience

Stage 1: Discrete Choice Experiment (DCE)
After the first set of demographics, each participant is administered ten discrete choice (or stated-preference) experiments. They have to choose the preferred job between two hypothetical job offers each time, presented in a table format. An additional attention check job comparison brings the total to eleven choices, as explained below.
Each job offer varies in terms of the following attributes: wage relative to their current wage (if employed) or to what they consider a fair wage for a role in their desired industry (if not employed) – same, +5%, -5%, +10%; gender of the direct supervisor – female, male; schedule flexibility arrangements (relative to a standard 40-hour working week) – no flexibility, 1 hour flexibility at beginning and end of workday, entirely flexible schedule; gender composition of immediate colleagues- similar share of females and males, majority females, majority males; skills development opportunities – low, moderate, high.
We instruct respondents to assume that any job attributes not mentioned in the table are identical across the two offers.

These attributes and respective levels yield us a factorial design size of 216 job profiles. Based on these 216 possible job profiles, we obtained an initial full factorial design made up of 23,220 (=216 x 215/2) possible job pairs. From this initial full factorial design we excluded the pairs containing strictly dominated jobs. To define dominance, the two jobs are compared in terms of wage, flexibility arrangements and skills development opportunities, keeping the other two gender attributes constant across the two. One job is hence dominated if it presents lower levels on one, two or all three attributes, keeping all other attributes constant. In addition, we also excluded the pairs where either the supervisor gender or the colleagues gender composition alone was the only attribute varying between the jobs, to avoid forcing participants when they may be indifferent on those characteristics. To account for this decision, we complement the survey asking direct stated preferences on these attributes, as explained in detail below.
In this way, the final clean factorial design containing all possible job pairs contains 20,952 possible pairs. Based on this, we create a row random design for our discrete choice experiment, where the 10 choice tasks are randomly selected without replacement from the clean factorial design, so that all 20,952 eligible pairs are shown at least once across the sample. We ensure this by subsequently excluding the 10 choices randomly chosen for each participant from the pool of remaining pairs from which to choose those for the following participant. This implies the existence of a “boundary” participant, for whom there will be 2 pairs left from the initial set, whereas the other 8 pairs will be randomly drawn once again from the whole clean factorial design, in a second cycle of random draws.
With this design all 20,952 possible job pairs will be presented once during the experiment, while an additional set of 4,048 twice.

In addition to the ten choices explained above, participants are shown one comparison serving as attention check, where the jobs differ only in terms of wage and flexibility. One of these jobs is clearly dominant (current wage and no flexibility vs wage +10% and fully flexible schedule). This pair is randomly shown between position 6 and 10, increasing the number of choices to eleven. The choice from this comparison will be excluded from the willingness to pay estimation.
After completing the eleven choices, participants are shown again the last job offers and their selected choice, and they are asked to provide a short motivation for their choice in an open-ended question.

Stage 2: Perceptions about Managerial Characteristics
The second stage of the experiment begins with participants rating a list of seven characteristics in terms of how important they consider them in a direct supervisor at work. The attributes are: effective leadership; fairness; empathy; work-life balance support; development opportunities; autonomy granted at work; and work culture. Each characteristic is rated on a 1 (not important at all) to 4 (very important) scale.
Next, for each characteristic respondents indicate whether they think a female or male supervisor would score higher on that characteristic. The attributes are presented using both a positive framing (e.g. “more objective and fair” for fairness) and negative framing (i.e. “less supportive of work-life balance” for work-life balance support), so that systematically choosing the same gender does not always imply a positive attribution to that gender. The positively framed characteristics are fairness, development opportunities and work culture. The negatively framed ones are effective leadership, empathy, work-life balance support and autonomy.
In presenting the question, we randomise: the order of female or male supervisor in the choices presentation of the question, to reduce framing effects by presenting one specific gender as leftmost option; the order of the characteristics.

Stage 3: Actual Job Preference
In the last stage participants are told that a real anonymous big company has an above-average share of either females or males in the top earning quartile (which proxies as likely manager gender). They have to hence state whether they would be interested in finding out more about job opportunities at such company. If their answer is affirmative, they are shown a link to the real company’s webpage of job vacancies. We track if they actually click on the link or not, to compare if stated preferences align to actual behaviour.

Additional parts of the survey
Finally, as additional section, we ask participants to state their preferences regarding the gender of their direct supervisor (female, male or no preference) and of their immediate colleagues (majority females, majority males, similar share of females and males, no preference). For each answer we ask them an open-ended question to motivate their choice as well.
This last section is complementary to the previously indicated decision to exclude from the discrete choice experiment any job pair where only one of the gender attributes varied. Indeed, in these stated preference question respondents have the possibility to indicate their indifference as well.
This section is administered after all experimental stages to avoid anchoring respondents’ choices in various stages.

We will check for inattention in 3 ways.
First of all, we use a traditional attention check question, in which we indicate specifically in the question text how participants should answer. Next, as already outlined above, we include one job comparison as attention check in the DCE, with one job strictly dominating the other. This will be positioned randomly between comparison 6 and the second to last one (10). Finally, a third attention check consists of an image with two arrows (a modified Muller-Lyer illusion) and a question asking which one is larger.
In addition to inattention, a second source of bias could be the use of AI for the completion of the survey. To flag AI usage, we employ three strategies as well. First, the third attention check represents also a so-called AI cognitive trap, as AI models will fail to answer correctly and indicate the traditional answer from the Muller-Lyer illusion (Affonso 2026). Next, we exploit the open-ended questions and transform them into honeypot questions, where a JavaScript code requires AI models and bots to answer in very specific ways, completely unrelated to the survey. However, human participants will be free to answer in the way they consider appropriate as the honeypot indications will be hidden to them. Finally, we will also employ the Prolific AI check. Any answer coming from respondents flagged as AI will be excluded.

Not available

For Stage 1: row random design at the respondent level – ten pairs of jobs (ten rows) randomly selected from the clean full factorial design for each respondent. In addition, one attention check pair, with a strictly dominant job, randomly included among the shown choices, from position 6 to 10 out of 11.

For Stage 2: randomisation of order of attributes in importance rating question; randomisation of order of attributes and of order of which is leftmost between female or male supervisor in perceptions about supervisor.

For Stage 3: within the industry respondents work in (if employed) or within the desired industry (if not employed), each respondent is randomly assigned to either a “female” or “male” company, meaning a company with either high share of females or high share of males in the top earning quartile.

Stage 1: respondent level (random assignment of choice sets and attention-check position)

Stage 2 and 3: respondent level

No

No clustering of the treatment

Target sample size: 2,500 participants recruited using Prolific platform. We will target a sample of individuals living in the UK, evenly split by gender. Individuals will be in the labour force, selected using the employment status filter provided by Prolific with the following categories: full-time employment; part-time employment; due to start a new job in the next month; unemployed (and job seeking). We exclude self-employed individuals from our sample.

Total sample size: N=2,500 respondents with complete responses

Stage 1: All respondents complete the job-choice experiment. Choice sets are randomly drawn without replacement from the clean factorial design; therefore, there are no fixed treatment arms, and all 2,500 respondents contribute to this stage.
Each choice task is randomly drawn without replacement from the set of 20,952 eligible job pairs, according to the row random design explained above in the experimental design section. All 20,952 pairs will be shown once and out of these, 4,408 pairs will be shown twice.

Stage 2: All respondents are exposed to randomised ordering/framing conditions. Randomisation is approximately balanced across respondents.

Stage 3: Respondents are randomly assigned to one of two company-profile conditions: “female” company arm, with approximately 1,250 respondents; and “male” company arm, with the remaining 1,250 respondents.

All power calculations assume a significance level of α=0.05 and a statistical power level of 1-β=0.80. For the main outcome 1: Minimum Sample Size for Willingness to pay (WTP) for manager gender. We assume a minimum detectable effect size of δ_FemSup=0.10 for the parameter corresponding to female supervisor in the main specification of WTP estimation (see pre-analysis plan below for estimation details). This would correspond to a WTP for a female supervisor of approximately 7% of current wage (using multinomial logit model) to 8% (using OLS). For comparison, the estimated WTP for gender diversity from Högn et al. (2026) is between 2 and 3%, the WTP to avoid risk of victimisation and harassment between 20 and 30% in Folke and Rickne (2022) and 37 to 50% in Curull-Sentis et al. (2025). To determine a minimum sample size for the WTP estimates we follow the approach of de Bekker-Grob et al. (2015) and use the following formula (corresponding to their eq. 4): N>[((z_(1-β)+z_(1-α) √(Var(δ_FemSup))))/δ_FemSup ]^2 The sample size is calculated using the standard normal quantile for 1-α and 1-β, the minimum detectable effect size δ_FemSup, and an estimated variance for δ_FemSup (for details on variance estimation for parameters of multinomial models, please refer to McFadden 1972). This results in a minimum sample size of 2,429 respondents which we round up to 2,500. For main outcome 2: Click on company website The 2,500 participants will be equally split across the two arms of the treatment (with approximately 1,250 respondents per treatment arm). With a sample of this size, it is possible to detect a minimum effect of 0.11. For main outcome 3: Attribution of positive managerial characteristics to male supervisor The sample of 2,500 participants allow to detect a minimum effect of f^2=0.003 for a two-sided test, using a linear multiple regression model with eight predictors and testing a single regression coefficient. This calculation should be interpreted as approximate, as the G*Power procedure does not account for the clustering of observations within respondents.