Habyarimana, J., Humphreys, M., Posner, D. N. & Weinstein, J. M. (2007). “Why Does Ethnic Diversity Undermine Public Goods Provision?” American Political Science Review 101(4): 709–725. DOI 10.1017/S0003055407070499. Only the PDF was available: no data, no code. Every number below is a reported estimate from Tables 1–9 or the text.
Primary: estimates. A field lab in Mulago-Kyebando, a poor, mixed neighbourhood of Kampala; 300 randomly sampled residents play a battery of games, each built to isolate one candidate mechanism, sometimes with co-ethnics and sometimes not. Table 1 (taste survey), Tables 2–3 (dictator game, giver anonymous), Table 4 (puzzle game), Table 5 (network game), Tables 6–8 (dictator game, giver seen), Table 9 (prisoners’ dilemma with and without an enforcer).
Secondary: an argument, the paper’s taxonomy of three families and five mechanisms (shared tastes, other-regarding preferences, efficacy, findability, strategy selection by sanctioning), and the logic of which game separates which mechanism from which.
Identification, with a counterintuitive core. Mixed communities provide fewer public goods, and the obvious story is that people care more about their own. The paper’s design shows they do not: when the giver is hidden, people give their own group no more than anyone else. What changes behaviour is being seen: the stingiest 40% give their own group much more once their face is shown, co-ethnics can find each other, and enforcers punish co-ethnic free-riders more often. Shared tastes and better teamwork find no (or only weak) support. The lesson the player should leave with: it is not kindness, it is being watched by people who can find you — and you only know that because one test was run with the giver hidden and again with the giver seen.
In a poor neighbourhood of Kampala, streets where neighbours share a background keep their drains clear and their nights patrolled; mixed streets often don’t. There are five possible reasons. Some researchers invited 300 neighbours to play money games to find out which is true. You are the researcher, but you can only afford five of the games. You’ll find that people are not kinder to their own when nobody is looking. Something else happens when they are seen.
Statistical, from reported estimates only. Nothing is simulated; there are no random draws. Each test shows the paper’s reported estimate and whether the paper found it distinguishable from chance.
Aim (player’s words): “Five reasons are suspected. Run five tests, then call each one — and back every call with a test.”
Verdict (end screen, top): “You proved k of 5 calls.” Then: “The study’s answer: people were not kinder to their own. When nobody could see them, they shared equally. Once they were seen, the stingiest gave their own group much more — and their own group could find them.” Then: “So mixing need not fail. What makes a street work is that people believe they are seen.”
Three concepts at different positions:
A. The Inquiry (Experimenter, challenge + discovery). Five suspect explanations, seven of the paper’s games as tests, budget of five. Tap a test, bet on its outcome, watch two bars race (own group vs other group). At the end, call each suspect Guilty / Not guilty / Can’t tell; a call scores only if it is right and backed by a test you ran. Deliberation: which test speaks to which suspect, and what pair of tests is needed to tell kindness from being watched. Fun: anticipation (bet, race), agency (budget allocation decides what you can prove), discovery (the hidden/seen contrast), mastery (second run with a better test plan). Replay: a different test plan (the prisoners’-dilemma route to the “watched” verdict instead of the dictator route) is a genuinely different solution.
B. The Giver (Subject/Agent). The player splits coins between two faces, sometimes seen, sometimes not, and is then shown how Kampalans split. Deliberation is real only about one’s own taste; the finding is about others; the player’s own play teaches nothing about the mechanism. Demonstration-heavy.
C. The Organiser (Controller). Run a mixed street; choose policies (appeal to generosity, team training, make contributions visible, split the street) and watch contributions. The cause–effect lever would extrapolate lab estimates to a street the paper never measured; illustrative at best, and the payoffs would be ours.
0:00 — Two little streets: one tidy with one-coloured houses, one with a clogged drain and mixed colours. “Why does the mixed street fail?” The aim line. I press Start. 0:20 — Five suspects at the top, each a short card: Different tastes, Kinder to their own, Better teamwork, Easier to find, Watched by their own. Seven test cards below, “Tests left: 5”. I’m weighing: kindness seems obvious; which test checks it? “Share the coins — giver hidden” sounds right. 0:50 — I tap it. Three portraits appear; the giver’s is a grey silhouette. “Will givers give their own group more?” I bet yes. Coins fall into two piles: own group 265, other 276. “No difference — if anything, less.” I frown: so not kindness? 1:30 — “Share the coins — giver seen”. Same, but the giver’s face is shown. Bars: most people still equal; the stingy 40% row jumps: own group 178 vs 142. Grin: they’re kind only when watched. 2:30 — I spend the last three on the survey, the puzzle and the network game. Puzzle: 48% vs 38%, “could be chance”. Network: 43% vs 28%, a real difference. 3:30 — Verdict sheet: I call each suspect. Kindness: Not guilty. Watched: Guilty. Puzzle: I hesitate; I call Can’t tell. 4:00 — Reveal: 5 of 5 proven. Had I skipped the hidden test, the sheet would have told me “Seen alone can’t tell kindness from being watched.” I tap Play again to try the other route.
The runner-up B fails at minute one: “I split coins evenly because that’s what I’d do” — nothing learned about Kampala.
A. It puts the player where the paper’s insight lives (the design of the comparison) and the discovery lands in the second or third tap. Runner-up B; switch if testing shows strangers cannot keep five suspects in mind (then cut to three families).
One round of five tests, one verdict sheet, one reveal. Screens: start, lab (suspects + tests), test sheet (bet → race → result), verdict sheet, end, how to play, about. No presets beyond the seven games, all the paper’s. No order is imposed; the hidden dictator game sits first in the list so the surprise can come early. Score: proven calls out of 5, only at the end; during play the only counter is tests left.
mechanism.js): the table of reported results; the evidence rules (which tests back which verdict, from the paper’s own use of each game); scoring.Region-based and subjective co-ethnicity codings (we use benchmark ethnic co-ethnicity; subjective given in the result note where it is the stronger figure); the 500 USh “discrimination” variant of the dictator game except as a footnote; lockbox game (mentioned in the puzzle result); individual group dummies of Table 1; enforcer behaviour by type; all regression controls.
From the paper’s Figure 1–2: the Public Information Box, three portraits in a row, the giver’s replaced by a grey silhouette when anonymous. Rules: two group colours only (teal = giver’s own group, ochre = other); grey silhouette means hidden, a face means seen; every result is the same two-bar race, bars at least 40 px tall; no more than five suspects and seven tests on any screen. Legibility: on the test sheet the player looks at three portraits (≈70 px) answering “who is seen?”, and two bars answering “who got more?”.
Headless (fidelity.js, Node): (1) every displayed gap equals the table value (exact); (2) “clear” flags match the paper’s significance stars; (3) derived type averages reproduce the reported overall averages within 2 USh; (4) bar pairs reproduce the reported gap exactly and their mean equals the base; (5) the evidence rules: the paper’s own route (survey, hidden DG, seen DG, puzzle, network) can prove 5/5; seen DG without hidden DG cannot back “Kinder to their own: not guilty” or “Watched: guilty”; (6) scoring: right-but-unbacked calls score 0; (7) every 5-test plan of the 21 possible is scored, reporting how many reach 5/5 (the discovery does not hinge on one ordering).
See the end of this file (appended after the build).
node fidelity.js, full output in fidelity-result.txt)ALL PASS. Every displayed gap equals the table value (Tables 2, 4, 5, 6, 9; p. 723); the “clear / weak / could be chance” tags match the paper’s stars and wording; derived per-type averages reproduce the reported overall averages (230 and 226 USh) exactly; bar pairs carry exactly the reported gap. Evidence rules: the paper’s own route proves 5/5; “Giver seen” without “Giver hidden” backs neither the altruism nor the sanctioning verdict; the prisoners’-dilemma route backs sanctioning but not altruism (4/5). Over all 21 five-test plans, scored with the paper’s calls: 1 plan proves 5, 12 prove 4, 8 prove 3. Results do not depend on test order. The central discovery (hidden gap ≈ 0, seen egoist gap +36 USh, enforcer erases the co-ethnic edge) is reported data and is shown identically whatever the budget or scoring.
mechanism.js written with all numbers in CONFIG; no random draws. Fidelity check passed.loop.js: two headless sessions (a first-timer who runs “Giver seen” first and convicts kindness, 3/5; the paper’s route, 5/5). Read as a stranger: each test changes something visible, the bet makes a commitment, and the wrong conviction is explained in the verdict.session-test.js in headless Chromium at 380 px: a full first session with help opened and closed mid-test, a retry along the paper’s route reaching 5/5, About opened and closed from the end; no console errors; no horizontal scroll on any screen (scrollWidth 380). Log in session-log.txt. Session length is an estimate.