Design

• Recruited from baseline, conditional on interest
• Task force: 10 appointed members (board, experts, community leaders) plus 5 randomly selected residents
• Control: interested residents who were not assigned

Invitation is random among those who said they wanted to come. Effects are for that willing set. The LATE is for those who attend at least two meetings.

Scheme · Timeline · Attendance · Who attended · Context

Timeline

Implementation ran over roughly two years in 23 municipalities.

• February 2024. Baseline begins (first 11 municipalities).
• Late 2024. Working groups begin, then suspend in September.
• September–November 2024. War; programming interrupted.
• Early 2025. Second baseline wave (13 further municipalities). Meetings resume.
• May 2025. First municipal elections in nearly a decade.
• After project selection. Endline survey (trust index and allocation).
• Interviews. 47 structured interviews; 46 enter the mixed-methods analysis (26 linked to the survey).
• From March 2026. Renewed conflict.

Context · Attendance

Interest

Stated interest is roughly bimodal. 28 percent say they are definitely not interested; 30 percent say they certainly are.

• Willing set somewhat more male and somewhat younger (\(p < 0.01\); age \(p < 0.01\))
• Those claiming corruption marginally more likely to volunteer (\(p < 0.05\))
• Those providing a larger allocation more likely to volunteer (\(p < 0.01\))
• Differences are not huge

Actual attendance

28 percent of baseline respondents say they do not want to take part. Among those invited, 37 percent attend no meetings. 43 percent attend at least two.

Attendance among randomly assigned participants was low. People had other things going on. Context

Estimated effects

ITT and LATE on allocation: Amount, More, and Some. Subgroup intervals are wide. No statistically significant heterogeneity by gender or local marginalization.

ITT · LATE · Trust items

Primary results: ITT

LATE · Estimated effects

Primary results: LATE

ITT · Estimated effects

Working graph

The graph represents a mid-level theory.

No leftover path from invitation or participation into trust. Paths run through inclusion, understanding, and connections. Full theory DAG

Completing types

The working graph implies a large type space. The unrestricted binary model has 8,796,093,022,208 types (\(2^{43}\)). Count

We restrict: an individual’s effect may be positive, negative, or flat, but may not change sign across backgrounds. A three-parent node then has 104 types rather than 256.

Interviews supply two things the survey does not:

1. A factual parent–child pair (scores cut at 0.5)
2. A single-parent attribution (the 4-by-2 uses cutoff 0.75)

Cells that would require flipping two or more parents are not elicited. The unit’s count is split across the remaining types. The survey has levels only.

Raw scores

Each arrow is a 4-by-2: factual pair by whether the coder attributed a difference.

Invitation \(\rightarrow\) participation: 1 positive, 2 negative, 43 none. Hard four-type counts: 21 chronic, 22 destined, 2 adverse, 1 beneficial.

Most invited participants do not attribute participation to the invitation. All arrows · Quotes · 4-by-2 note

Quotes: \(n_{+}\), \(n_{-}\), \(n_{0}\)

Curated excerpts from the paper quote table. One excerpt per sign where one exists. No unit identifiers.

All arrows · Raw scores

Invitation to trust

Query mean, with query standard deviation in parentheses. The columns are not four posteriors from the same model.

Share rows · Validation

Agree and disagree

They agree on the major moving parts.

• Inclusion, understanding, and connections raise trust on both sides
• Marginalization lowers connections and inclusion
• Participation raises connections (0.062 interviews; 0.039 survey)

They disagree on what the treatment did.

• Invitation \(\rightarrow\) participation is small on both sides and weaker in the interviews (0.007 vs 0.035)
• Participation \(\rightarrow\) inclusion and participation \(\rightarrow\) understanding are positive in the interviews and approximately zero or slightly negative in the survey

The invitation-to-trust average is approximately zero on every column. The experimental allocation ITT is not on this graph.

Validation

26 of 46 interviews link to the survey. Small \(n\). Invitation agrees on 23 of 26 (0.88). Other nodes agree on about 0.65. Omnibus \(p = 0.028\).

Correlations are far from unity. No unit identifiers. Queries

Takeaways

• Demand for participation is low. 28 percent say they do not want to take part. Among those invited, 57 percent do not attend regularly.
• The trust index does not move. The allocation does, mainly at the low end.
• Qual and quant agree on the major moving parts. They disagree on what the invitation did.
• The working graph has no leftover path from invitation into a final outcome. It cannot speak to the experimental contrast as currently drawn.
• If a municipality already plans to include people, tihs is good news. But it does not provide grounds to set up or orient programs towards this goal.

Questions

Questions

1. Why did the allocation move, and mainly at the low end, while the trust index did not?
2. Are the two outcomes different objects, or is one just noisier? Linked-interview correlations sit around 0.3. Is that enough to treat the nodes as the same?
3. The working graph has one trust node and no leftover path from invitation. What graph would let the mixed-methods analysis speak to the experimental contrast?
4. Invitation barely moves participation in the interviews. Is that a first-stage failure, or a mismatch among three participation measures: LATE attendance of at least two meetings, the survey civic composite, and the interview coder score?
5. The 4-by-2 at 0.75 is weakly negative. The parent-effect query is weakly positive. Which object should update beliefs about invitation?

Back to headline · Extras

Context

Implementation ran through conflict, displacement, and the first municipal elections in nearly a decade (May 2025).

• Tensions from October 2023. War September–November 2024, and again from March 2026.
• Working groups suspended in September 2024 and resumed in early 2025.
• Some activities adapted (for example psychosocial support online).
• Newly elected officials sometimes shifted priorities away from earlier working-group lists.

Attendance was low. People had other things going on.

Back to design · Timeline · Attendance

Trust-index items

ITT and LATE on the index and its three items. Subgroup intervals are wide.

Back to trust · Estimated effects

Parent-effect numbers

Back to XY

4-by-2 versus the query

The Invitation–Participation 4-by-2 is weakly negative at cutoff 0.75. The parent-effect query on that edge is weakly positive (0.007).

A crossed-out cell pins the realized (marginalization, municipality) background only. Leftover types that are flat there can still move participation on the other backgrounds. The query averages over all four backgrounds.

Factual scores and the Dirichlet update cut at 0.5. The displayed 4-by-2 uses 0.75. Hard four-type counts then diverge: 1 beneficial type at 0.75, 5 at 0.5.

Back to raw scores · Back to questions

Unrestricted type count

Unrestricted binary types: \(2^{2^k}\) per node. Product across nodes: 8,796,093,022,208 (\(2^{43}\)).

The restriction (no sign change across backgrounds) is what makes the working model queryable. It is still a simplification: no unobserved confounding, and at most three parents per child.

Back to types

Share rows

Back to queries · Validation

Node means

Raw survey means of DAG nodes by assignment. Difference is treated minus control. These are descriptive, not query estimates.

Back to working graph