Using campaign platforms to measure local party ideology in Brazil

How does party ideology vary subnationally in Brazil? Although the country is known for its weak party system composed of many heterogeneous parties, there are still no suitable measures of party ideology at the local level. This project attempts to fill this gap by employing supervised machine learning methods. By cross-referencing roll-call voting ideal point estimation with campaign platforms of mayoral candidates and exploiting the fact that many legislators run for mayor, I build a model to infer candidate ideology from the text of their platforms. I then validate the model, applying it to candidates who ran in the 2020 São Paulo election and to all candidates of two main parties. Results are promising but still not super convincing: the model infers the fragmented MDB to be more cohesive than the disciplined PT. However, it does a good job distinguishing progressive and conservative candidates in the São Paulo election It needs refinements but is a first step in coming up with a convincing measure of local party ideology in a context where it is much needed.