On-the-Job Training for Politicians: Evidence from Brazil

The intervention provided to treated participants is Gabinete Compartilhado (Shared Cabinet), a continuous technical support program targeting first-term women city councilors in Brazil. The program offers a package of resources which include:
1. Legislative assistance: legal advisory in drafting legislation
2. Budget monitoring tools: courses and a hotline for urgent support on the municipal budget
3. Policy briefs on themes relevant to local legislative activity
4. Ongoing peer learning opportunities in courses and one-to-one meetings
5. On-demand meetings with specialists
Structured workshops and courses on evidence-based policies, political articulation, strategic planning and mental health.
Additionally, participants are encouraged to join Wave of Bills, a project that provides bill proposals based on scientific evidence and previously approved legislation, and reviewed by experts in the field. The bill proposals are focused on a myriad of public policy topics of general interest and relevance to their mandate, such as domestic violence and early childhood care. Finally, the NGO also offers legal support for cases of political violence.

2025-04-21

2028-12-31

The trial focuses on three main dimensions of impact: technical improvement of legislative activity, soft skills development, and political ability. The key outcomes are:
1) Legislative Output
a) self-reported ability to draft and submit bills
b) number and type bills proposed, approval rate of bills
c) share of policy-related draft bills proposed (instead of “symbolic”ones)
2) Executive Oversight
a) self-reported ability to conduct executive oversight activities
b) self-reported ability to interfere in the Municipal Budget
c) self-reported ability to participate in legislative committees
3) Strategic Planning and Mandate Management
a) structure of mandate activities and programmatic prioritization from questionnaire
4) Future in Politics
a) self-reported ambition in politics from questionnaire

The outcome variables from the surveys are constructed by a scoring (1-5) given by the interviewer team to each of the dimensions evaluated in the questionnaire immediately after the interview is done. The measures of legislative activity using data from the Chambers’ websites are constructed by evaluating the number of times the candidate participated in committees, proposed legislation, and convened public hearings. The quality of the legislation proposed is evaluated using a framework developed by the NGO to assess the policy-relevance of draft bills, mixing quantitative and qualitative analysis tools. Therefore:
1) Legislative output
a) scoring on questionnaire dimension (1-5)
b) approval rate = total number of bills approved / total number of bills proposed
c) policy-relevance = a measure constructed based on the framework developed by the NGO to analyze the text of draft bills
2) Executive Oversight: all variables derive from the scoring given in the relevant questionnaire dimension
3) Strategic Planning and Mandate Management: all variables derive from the scoring given in the relevant questionnaire dimension
4) Future in Politics: all variables derive from the scoring given in the relevant questionnaire dimension

Secondary outcomes include:
1) Mental Health and Team Management
a) questionnaire scoring
2) Occurrence and Vulnerability to Gender and Race-based Political Violence
a) questionnaire scoring
3) Political alliances/isolation within and outside of Chamber
a) questionnaire scoring
4) Party relationship and position within the party
a) questionnaire scoring

Mental health and political alliance and isolation measures reflect both the self-reported characteristics of city councils as well as perceived by interviewers, who then assign the scoring from 1-5. The team of interviewers is trained to assess qualitative aspects of the legislative activity and incorporate that into the scoring. An example is a candidate that states they do not suffer political violence cases but narrates cases which are considered by the interviewers to be political violence.

The study employs a randomized controlled trial design involving 83 first-term women city councilors across Brazil. Only one councilor per municipality is included in the study to avoid spillovers and competition, and stratified randomization is conducted based on quotas related to region, political alignment (left vs.center and right), and race (Black or Indigenous vs. other). Candidates are grouped into smaller groups of 3 to 15 individuals based on a PCA score defined by local and individual characteristics. One councilor is randomly selected within each subgroup to receive the treatment, while others are controls. Finally, subgroups are assigned to treatment or waitlist randomly. All estimates are weighted by the propensity score, which is calculated accounting for group-level quotas and stratification weights. Waitlist procedures are used to replace non-takers in a way that preserves the balance of the sample and the group-quota assignment.

Not available

The randomization is conducted in STATA, using cleaned and treated data from the Supreme Electoral Court on candidates of the 2024 plea, excluding those which appear as elected previously to 2024 to focus on first-term mandates. The waitlist is constructed such that we have enough replacements for at least 3 non-takers in each group. Replacement procedures rely on a pre-established waitlist ranked by subgroup scores, ensuring balance is maintained in case of attrition.

The unit of randomization is the individual councilor. To avoid contamination or spillover effects, only one councilor is selected per municipality to be potentially treated. The first step involved a within-municipality randomization that randomly allocated a rank to each female candidate in that municipality. Then, municipalities are chosen randomly to enter the sample, and the first-ranked legislator represents that municipality in the sample.

Yes

434 councilwomen, where 83 are treated and the rest are controls.

434 councilwomen.

83 councilors treated, and 351 in control.

Our study sample was pre-defined to include only women on their first mandate, to focus on politically inexperienced candidates. Due to limited availability of CPF (tax code) data in 2024, we decided to take reelection rates in 2020 for women elected for the first time in 2016, to avoid errors in classification related to the missing tax code. Using reelection attempt and success rates between 2016 and 2020, we obtain: Unit and outcomes: The unit of analysis is the candidate. The main outcomes are binary: (i) attempted reelection in 2020 and (ii) won reelection in 2020, defined for the cohort of women serving a first mandate in 2016 (N = 4,779). Baseline rates and standard deviations: The baseline rate of attempting reelection is 0.597, with a standard deviation of 0.491. The baseline rate of winning reelection is 0.306, with a standard deviation of 0.461. Sample design and clustering: Randomization is at the candidate level with stratification. Analyses will include strata fixed effects and cluster-robust standard errors at the stratification (subgroup) level. Power is computed using the realized sample sizes (83 treated, 348 controls), which reflect the stratified design and incorporate the effective group sizes after clustering. Minimum detectable effect sizes (alpha = 0.05, power = 0.80): For attempting reelection, the minimum detectable difference is 15.7 percentage points (from 59.7 percent to 75.3 percent). For winning reelection, the minimum detectable difference is 16.7 percentage points (from 30.6 percent to 47.3 percent). This means the study is powered to detect effects of roughly 16–17 percentage points on the main reelection outcomes.