How Elastic are Consumption Stereotypes?

We conduct two different experiments randomly assigned to two samples of participants.

The first experiment is designed as a gender identity prime. The gender identity prime is predicted to increase adherence to one’s own gender identity, and consequently increase adherence to the specific stereotypes of this particular identity (Benjamin et al. 2010, Benjamin et al. 2016, Coffman 2014). In other words, we expect this prime to increase the gender gap in consumption as captured by our primary outcome variables, without affecting the implicit association between products and masculinity.
The gender identity prime we use is inspired by a wide body of research studying two axes, typically labeled communion and agency (see Eagly et al. 2020 for a review and meta-analysis). Communion, prevailing in the female stereotype, is a set of traits that orient people to others and their well-being. Agency, prevailing in the male stereotype, is a set of traits that orients people to their own mastery and goals. A third set of traits, referred to as competence, has been also studied in gender stereotypes and is typically found more neutral in current surveys.
We follow Winterich et al (2009) to prime gender identity through the communal or agency themes. This prime involves a small essay focusing on agentic or communal qualities. Participants are then asked to write a paragraph discussing how they embody those qualities in their life. Each participant is randomly allocated between the agency treatment, the communal treatment, and two controls described below.
To write the small essays, we obtain traits frequently used in surveys using the meta-analysis across surveys of Eagly et al. (2020). We take the three most frequent traits in each axis across surveys since the 2000s. We obtain emotional, compassionate and affectionate for communion, and ambitious, assertive and decisive for agency. Those traits were further validated as being perceived by the average US respondent as mostly true of women (men) in the pilot study. We then wrote a small essay for each axis similar in length and structure on how these three traits may impact one’s life.
As a manipulation check, we ask participants to rank themselves on a continuous scale from feminine to masculine after the treatment and consumption choices (Brenøe et al. 2022).
Our active control is to propose the same exercise but on a third axis than communion or agency, referred to as competence. This axis typically appears more gender neutral in recent surveys (Eagly et al. 2020). The three most frequent adjectives in this axis across surveys are: creative, intelligent and innovative.
Our mock control is to complete an unrelated task in order to make the total survey duration identical for the mock control and treatment groups. The information of the mock control group should be neutral concerning the issues of interest (gender and consumption), which is typically difficult to find (Stantcheva 2022). We ask participants to imagine that they have the ability to time travel (past and future), and make them read a small essay similar in length around the potential places and people to witness with time travel. They then have to write a paragraph on a place they would want to explore if they had this ability.

The stereotype awareness experiment is an Implicit Association Test (IAT) predicted to reveal the automatic and unconscious association between gender and red meat consumption (versus fruits and vegetables). The Implicit Association Test (IAT) is a psychological tool originally used to measure implicit biases or unconscious associations between concepts. In the context of gender stereotypes, the IAT can reveal implicit associations individuals have between gender and specific qualities or roles, but also specific products or brands.
While the primary purpose of the Implicit Association Test (IAT) is to measure implicit biases, research suggests that increasing awareness of implicit biases through the IAT itself, or feedback from the IAT results, can potentially lead to changes in subsequent behavior. The process of becoming aware of one's implicit biases (i.e. using the IAT as treatment) can prompt individuals to reflect on and challenge their own stereotypes and biases, without priming/threatening their identity. In fact, the IAT is used in studies to (successfully) debias participants (Alesina et al. 2018, Gino and Coffman 2021).
In a pilot study randomizing the order of an IAT task before (vs. after) a module asking respondents to report their consumption, we found evidence that men (not women) “treated” with the IAT before the consumption module (vs. after) reported less consumption of red meat. This further motivated the use of the IAT-as-treatment in the pre-registered study. We expect this exercise to decrease the gender gap in consumption.
In our main treatment IAT, participants are involved in a computer-based task where they are presented with male and female names, and images of meat products or fruits and vegetables products. They are then required to categorize these stimuli by pressing specific keys on the keyboard. The test measures how quickly participants associate male names with meat-related pictures and female names with fruits and vegetables, compared to the reverse association. The IAT assumes that if an individual's response is faster for one combination (e.g., male-meat, female-fruits and vegetables) compared to the other (e.g., male-fruits and vegetables, female-meat), it suggests a stronger association between it. After being treated, respondents are provided with a short feedback in which we inform participants that the test they completed measures the automatic connections and biases we have between different things without consciously thinking about them, and ask them to report whether they noticed any automatic connection in their case.
Control participants are randomly allocated to one of three control IATs:
Active control I - gender axis: IAT asking to classify men (versus women) names and dice dots between 1-3 and 4-6. We expect no association between the two categories.
Active control II - food axis: IAT asking to classify common short (versus long) names (less than 4 letters versus more than 8 letters) and red meat products (versus fruits and vegetables). We expect no association between the two categories.
Mock control: IAT asking to classify common short (versus long) names and dice dots between 1-3 and 4-6. We expect an association (short names and low number of dots), but not connected to the gender-food axes.

2023-06-05

2023-06-30

Our primary outcome variables are intended at capturing respondents’ food consumption preferences, which are elicited in two ways (direct vs. indirect):

- Direct elicitation:
Participants are asked whether they intend to adjust their consumption of (i) vegetables and fruits, (ii) beef, pork and other red meat, and (iii) chicken, fish and eggs by increasing, maintaining or decreasing their future consumption.
They are also asked to think about a pretend trip to the grocery store, and to allocate $100 between 4 categories: (i) beef, pork (such as bacon), or other red meat; (ii) chicken (or turkey), fish (such as tuna, salmon, or shrimp), and eggs; (iii) fruits (like orange juice, fresh apples, or grapes) and vegetables (such as tomatoes, potatoes, or canned beans); (iv) other food items (such as dairy products, baked goods, or cereals).


- Indirect elicitation:
Participants are asked to complete a choice-based conjoint study involving choosing frozen pizzas. In each round, participants are presented with two frozen pizza products, both featuring tomato sauce and Mozzarella cheese, but with varying attribute levels on 5 major attributes (brand, crust, price, nutrition, and extra toppings). There is a total of 15 rounds per participants and each pair of frozen pizza is generated by randomly selecting attribute levels. The conjoint study allows us to capture the importance of each attribute, measure the part-worth utilities respondents get from red meat (pepperoni, three meats toppings), white meat (chicken), or non-meat toppings (four cheeses, veggies, plant-based meat), estimate the price elasticity of red meat, and simulate market shares for red meat pizzas between treated and control respondents.

From the choice-based conjoint task we will estimate part-worth utilities for red meat (pepperoni, three meats toppings) and price elasticity using a logit model.

Our secondary outcome variables are intended at capturing whether our treatment affected:
(1) respondents’ self-views about their gender identity: we ask respondents to put themselves on a continuous single-item scale going from “very feminine” to “very masculine”.
(2) respondents’ own stereotypical association between red meat and masculinity: we ask respondents to what extent they would perceive the following 3 categories as feminine, masculine, or neither on a continuous single-item scale going from “very feminine” to “very masculine”: "Fruits and vegetables", "Beef, pork and other red meats", and "Chicken, fish and eggs".

Besides allowing us to test for the existence of any “first stage” of our treatment, those variables can further be used to motivate the main channels through which our treatment might affect red meat consumption for men (vs. women).

We then ask a set of questions intended at capturing alternative potential mechanisms through which our treatments may affect food preferences across gender, namely:
- Respondents’ self-views about whether red meat tastes better, about whether they think they need to eat red meat every week, and their self-assessed ability to cook red meat better than vegetables. Those variables were strongly correlated to meat consumption and shown to explain a significant fraction of the gender gap in the pilot study.
- Respondents’ health, environment and ethical concerns, and whether they think red meat is bad for the environment, unhealthy or unethical. Those variables, although correlated to meat consumption, were shown not to explain the gender gap in the pilot study.

We also end the survey with socio-demographics questions and other control variables that we can use to explore potential treatment effect heterogeneity, namely respondents’ income, age, employment status, level of education, race, marital status, family size, political affiliation, TV exposure, and measures of health and physical conditions (weight, height, and physical activity).

The pre-registered online survey experiment is aimed at a sample size of about 4500 American participants. It is intended to be representative of American consumers above 18 using gender, age and US states quotas. Survey participants are collected via Qualtrics.

After committing to provide truthful answers and accepting to participate in the experiment by agreeing to a short consent form, participants are asked about their age, ascribed gender at birth, and US state of residence. They are then randomly allocated to one of 8 treatment arms (cf. intervention):
a. Prime condition: treatment agency prime, treatment communion prime, active control competence prime, mock control time travel.
b. IAT condition: treatment IAT, two active control IATs, one mock control IAT.

After being treated, we elicit respondents’ consumption preferences (cf. primary outcome variables), their self-views about their gender identity, and their stereotypical association between food and masculinity (cf. secondary outcome variables).

We then ask questions capturing more general beliefs about red meat (taste, needs, cooking skills), along with health, environmental and ethical concerns. Finally, we conclude the survey by gathering socio-demographics and health-related information on our respondents. We also include a feedback box to gather any comments participants may have about the study.

The online survey experiment was informed by a pilot study of 1000 participants conducted by the authors. The pilot study replicated the existence of a red meat gap documented in alternative sources (CEX, Nielsen), allowed the authors to explore potential factors correlated with this gender gap, and test for the robustness of questions that we used in the pre-registered online survey experiment.

The randomization of treatment arms across individuals is done by Qualtrics using quotas by age, gender and US States to be representative of American consumers above 18.

Randomization at the individual level.

No

4500 respondents

4500 respondents

a. Prime condition: 50% of the 4500 respondents (2250 respondents)
i. 1/3 receive the treatment “agency” prime (i.e., 750 respondents)
ii. 1/3 receive the treatment “communion” prime (i.e., 750 respondents)
iii. 1/3 receive the mock control (“time travel”) or the active control “competence” prime (i.e., 375 respondents per control condition)

b. IAT condition: 50% of the 4500 respondents (2250 respondents)
i. 50% receive the treatment IAT red meat/fruits veg (i.e., 25% of all respondents, or 1125 respondents)
ii. 50% receive one of the 3 control IATs (i.e., 375 respondents per control IAT)