The Effect of Gendered Language on Perceived Gender Suitability of Occupations: Evidence from an Experiment with Middle School Students

Participants are presented with a list of 25 occupations and sports and asked to indicate whether each is more suitable for men, for women, or for both. The list includes STEM occupations (e.g., engineer, biologist), non-STEM occupations (e.g., lawyer, photographer), and sports (e.g., skiing, cycling). Items are further classified as grammatically declinable — i.e., occupational titles that have distinct masculine and feminine forms in Italian (e.g., ingegnere/ingegnera) — or grammatically invariable — i.e., titles whose form does not vary by gender (e.g., farmacista). Participants are randomly assigned at the individual level to one of three language conditions in which the items are presented: in Condition 1 (generic masculine), all items are presented using masculine grammatical forms; in Condition 2 (pair form), declinable items are presented with both masculine and feminine forms; in Condition 3 (professional domain), items are replaced by their corresponding disciplinary or professional domain labels. To illustrate how the three language conditions work, consider the item "Engineer": in Condition 1 it appears as ingegnere (generic masculine), in Condition 2 as ingegnere/ingegnera (pair form), and in Condition 3 as ingegneria (professional domain). For invariable items such as "Data analyst", the form is identical across Conditions 1 and 2 (analista dati), while Condition 3 uses the domain label (analisi dati).

2026-05-20

2026-06-19

A binary indicator (1 = suitable for both genders, 0 otherwise) for each item × student observation, estimated separately for boys and girls. The main model uses the sample of grammatically declinable items (19 items) for hypotheses H1, H2,H3, and H5. A separate model on grammatically invariable items (6 items) tests hypothesis H4.

For each occupation and sport in the list, participants indicate whether they consider it more suitable for men, for women, or for both. The outcome is a dummy variable equal to 1 if the participant responded "both" for a given item, and 0 otherwise. The estimand is therefore a vector of binary responses per student, estimated separately for boys and girls. Higher values indicate more egalitarian perceptions.
Item categories are captured by a categorical variable with three mutually exclusive values — STEM, non-STEM, and Sport — coded as two dummies with non-STEM as the reference category (D_STEM and D_Sport). The language condition is coded as two dummies (D_pair and D_domain) with generic masculine as the reference category.
Hypotheses H1, H2, H3, and H5 (see below) are tested in the main regression on declinable items. Hypothesis H4 is tested in a separate regression on invariable items only.

The secondary outcome examines whether the effect of linguistic framing on gender suitability perceptions differs between students who participated in the SPARKLE education program and those who did not.

We examine whether participation in SPARKLE, an education program evaluated through an RCT aimed at introducing middle school students to STEM through hands-on laboratory activities and at building awareness of gender stereotypes and self-confidence, moderates the effect of language condition on the primary outcome. This analysis uses the full sample (SPARKLE treated and control students). This analysis is exploratory, given uncertainty about the magnitude of the shift in gender stereotype levels produced by the SPARKLE intervention. The analysis is operationalized as two interaction terms between language condition dummies (D_pair and D_domain) and SPARKLE treatment assignment in an OLS regression, estimated separately for boys and girls. These estimands are complementary to — and do not overlap with — the main SPARKLE pre-registration, which estimates the main effect of SPARKLE on gender suitability perceptions, among other outcomes (for details, see the pre-registration titled "SPARKLE - STEM Practical Activities to Raise Knowledge Learning and Exploration: a cluster-randomised evaluation of a multi-component STEM, gender, and AI educational intervention in Italian middle schools. "); here, the object of interest is the interaction between SPARKLE and linguistic framing.

This is a three-arm between-subjects experiment embedded in the follow-up questionnaire of the SPARKLE RCT. The primary analytic sample is restricted to the control group of the SPARKLE RCT to obtain an uncontaminated estimate of the effect of linguistic framing. Students are randomly assigned at the individual level to one of three language conditions for a survey question presenting a list of 25 occupations and sports and asking whether each is more suitable for men, for women, or for both. The three conditions differ only in how the occupational items are linguistically framed: generic masculine job title, pair form (both masculine and feminine), or the profession name. Randomization to language condition is independent of prior assignment to the SPARKLE treatment.

Item list

The experiment includes 25 items across three mutually exclusive categories: 10 STEM occupations, 10 non-STEM occupations, and 5 sports. Within each category, items are classified as grammatically declinable (19 items) or grammatically invariable (6 items) in Italian.
- STEM occupations (10):
- Declinable in Italian (8): Computer programmer, Engineer, Architect, Biologist, Astronomer, Researcher, Surgeon, Nurse
- Invariable in Italian (2): Data analyst, Pharmacist

- Non-STEM occupations (10):
- Declinable in Italian (7): Psychologist, Police officer, Lawyer, Photographer, Pastry chef, Translator, Farmer
- Invariable in Italian (3): School principal, Social worker, Journalist

- Sports (5):
- Declinable in Italian (4): Skier, Football player, Dancer, climber
- Invariable in Italian (1): Cyclist

Hypotheses:

H1: Participants in the pair form condition will show more egalitarian perceptions than participants in the generic masculine condition (positive coefficient on D_pair).

H2: Participants in the professional domain condition will show intermediate perceptions between the generic masculine and pair form conditions (positive coefficient on D_domain, smaller than D_pair).

H3: The effect of language condition will be stronger for STEM occupations and sports than for non-STEM occupations, tested as positive interaction terms between language condition dummies and D_STEM and D_Sport.

H4: In the pair form condition, inclusive priming will extend to grammatically invariable items, increasing the probability of a “both” response relative to the generic masculine condition. Tested in a separate regression on invariable items only.

H5: The effect of language condition will be non-linearly moderated by pre-existing gender stereotype levels, measured at baseline prior to both SPARKLE treatment assignment and language randomization. Specifically, we hypothesize an inverted U-shaped relationship: the effect of inclusive language conditions (pair form and professional domain) relative to the generic masculine will be strongest for students with intermediate stereotype levels, and attenuated for students at both ends of the stereotype distribution. Students with very low stereotype levels (highly egalitarian) are expected to show a ceiling effect; students with very strong (negative) stereotype levels are expected to show resistance to inclusive linguistic cues. This is tested by including both a linear and a quadratic interaction term between language condition dummies and the baseline stereotype scale. The non-linear moderation hypothesis predicts a significant quadratic interaction term, with the effect of language condition being strongest for students with intermediate stereotype levels. Given the much lower levels of stereotypes among female students at baseline, the U shape may not be observed.

Estimation strategy:

Primary analyses (H1, H2, and H5) are conducted on the control group of the SPARKLE RCT only, using the sample of grammatically declinable items (19 items). The outcome is a binary dummy equal to 1 if the participant responded "both" for a given item. The model is an OLS regression with standard errors clustered at the individual level, estimated separately for boys and girls. All models include language condition dummies (D_pair and D_domain, with generic masculine as reference), item category dummies (D_STEM and D_Sport, with non-STEM as reference), and individual-level controls (e.g., cognitive ability, math grade, year of birth, parental education, language spoken at home). H5 adds the baseline stereotype scale, its square, and their interactions with language condition dummies. A separate primary analysis (H4) uses the same model structure estimated on the 6 grammatically invariable items. The exploratory secondary analysis uses the full sample (treated and control), adding SPARKLE treatment assignment and the interaction terms D_pair × SPARKLE and D_domain × SPARKLE.

Individual-level randomization using a computer-generated random assignment algorithm (python within oTree) at the time of questionnaire administration.

Individual student

No

Standard errors are clustered at the individual level to account for the multiple observations per student (one per item). 785 students (assuming no absenteism during the day of the survey).

19 (professions) x 785 (pupils) = 14,915

Condition 1 (generic masculine): approx. 261
Condition 2 (pair form): approx. 261
Condition 3 (professional domain): approx. 261

Based on results from the baseline (which included only Condition 3) and had a different list of professions (with some overlap), we estimate a within individual ICC of 0.28. This implies an MDES with no covariates of 0.14 SD which based on the baseline data SD, corresponds to 6.9 percentage points. Including covariates, which are expected to be highly predictive of outcomes (assuming an R2 of 0.5), the MDES is lowered to roughly 0.10SD or 5 percentage points.