🔍 Conclusion:
We disentangle taste-based and statistical discrimination in voting.
Despite gender identity being salient, biased beliefs — especially about policy — drive underrepresentation.
→ Correcting these beliefs can significantly raise female vote shares.

📈 Optimized messaging (tailored by municipality) can increase female vote shares by ~1.5 percentage points — at low costs. 20% municipios experience more than 2 percentage point increase.

Superb out-of-sample fit for
1) Full-Brazil without RCT municipios
2) RCT municipios' neighbors
3) 80-20 split out-of-sample validation

📊 Key Finding #6:
👉 Instead, underrepresentation is driven by statistical discrimination:
🔹 118% of the gap is explained by voters’ biased beliefs (about ability and especially policy positions)
🔹 –18% is attributed to taste-based factors

📊 Key Finding #5:
👉 Gender identity is very salient in voting decisions. Yet it does not explain women's underrepresentation. Why?
Because both men and women exhibit in-group preferences, and the net effect cancels out.

📊 Key Finding #4:
👉 Ability-based messages were less effective.

📊 Key Finding #3:
👉 Informative policy messages helped realign these beliefs and increased female vote shares.

📊 Key Finding #2:
👉 Many female voters believe male candidates are closer to their policy preferences than female ones — reflecting a disconnect between descriptive and substantive representation.

📊 Key Finding #1:
👉 Targeting male voters with gender identity messages reduces their distaste for voting against their gender — weakening taste-based discrimination.

🎯 RCT Design:
Large-scale digital campaign via Instagram, randomizing municipalities into seven groups:
Info Ability messages
Uninfo Ability messages
Info Policy messages
Uninfo Policy messages
Gender Identity messages targeted to men
Gender Identity messages targeted to women
Control

Voters don’t observe ability or policy — they form beliefs about them. Meanwhile, gender identity is observed and can be weighted more or less heavily (salience). The model separately identifies:
✅ How voters weigh each dimension (salience)
✅ What they believe about each dimension (expectations)

Model:
Voters choose candidates based on three dimensions:
• Gender identity (horizontal; identity → taste-based)
• Ability (vertical; beliefs → statistical discrimination)
• Policy alignment (horizontal; beliefs → statistical discrimination)

Why are women underrepresented in politics? Taste-Based or Statistical Discrimination?

We combine a structural voting model with a Randomized Controlled Trial (RCT) across 1,000 Brazilian municipalities during the 2024 local elections to find out.