Photo credits
  • Moneta Sleet Jr. - public domain, Saint Louis Art Museum
  • Heinrich-Boell-Stiftung - CC BY-SA 2.0
  • Kaizenify - CC BY-SA 4.0

Causal Presidents · the design-level results

The design-level results

The three figures that decided the verdicts below. Each is drawn by R from the same rows this page counts.

The figure is the artifact. If it were one re-fit behind the table, its recorded byte count would no longer match the file.

What was published

Forest plot of the published estimates, one panel per design, with each decisive check drawn hollow
two panels side by side, one per design, each with its own axis. One row per published specification: the estimate as a filled point with its 95 % interval, and the zero rule. The decisive check of each design is drawn hollow beside its estimate (A: the 24-month mis-timed dose that returns an estimate as large as the published one; B: the 6-month lead placebo), because a forest plot of estimates alone would show a number whose own artifact says it is not identified without showing why.

State dose-response DiD

Verdict estimated_not_identified

the lead-shift p-value is 0.167: 10 of 60 mis-timed doses return an estimate at least as large as the estimate itself

The checks that decided it

National VECM / ARDL time series

Verdict estimated

What the grid found

Specification curve: every fitted specification sorted by coefficient, with the axis levels that produced it
two panels (A above, B below), each sorted by coefficient. Top: estimate with its 95 % interval, coloured by sign, the zero line, and the pre-registered specification marked. Bottom: the axis levels of each specification in the same order, so a sign flip can be read against the choice that caused it.

60 of 81 specifications are positive, 21 negative, and 2 sit inside the near-zero band.

36 of 48 specifications are positive, 12 negative, and 22 sit inside the near-zero band.

Across the grid the coefficient runs from 0.3968 to 2.3976.

The in-time placebo returns p = 0.0625 over 48 shifted alignments.

The sign of Design B disagrees with the preferred one in 12 of 48 grid specifications exclude zero, but the in-time placebo does not clear 0.05: 3 of 48 shifted alignments return an estimate at least as large as the true one (p = 0.0625). The sign is the robust claim here; the timing is the weaker one, and it is what the placebo tests.

The synthetic counterfactual

Synthetic control (cross-country)

Verdict estimated_not_identified

9 outcomes, 31 windows estimated against a synthetic Nigeria built from 99 donor countries; 14 windows refused by the declared rule, and 0 survived their placebos. Each outcome is its own fit in its own unit, so a gap of 2 in one is not twice a gap of 1 in another. The bracketed figure is the same problem under a ridge refit: where the two disagree in sign, the estimate belongs to the solver rather than to the data.

One outcome was fitted a second way in 2026, with an estimator that does not depend on the 99-donor simplex: under-5 mortality. It refused as well, every window sitting inside the range that treating a random country produces, so the refusal does not rest on the ridge check alone. That second fit is reported in the project's documentation, not on this page, and no number from it is published.

Under-5 mortality rate (per 1,000 live births)

Maternal mortality ratio (per 100,000 births)

Penta3 immunisation coverage (% surviving infants)

Skilled birth attendance (% of births)

Every declared window was refused for this outcome: Nigeria's series cannot fill a window of this design's declared length. The card carries the reason.

Reported security fatalities (count)

Reported security incidents (count)

Real GDP growth (y/y) (%)

Headline inflation (y/y) (%)

Unemployment rate (%)

A synthetic control has no sampling distribution, so there is no interval here and none of these gaps is a card number. An outcome whose every window was refused is listed with the reason instead.

Did anything move at a transition?

Four transitions, two outcomes, and a placebo drawn from every eligible month. The rows below name the month, not the person: a shift at a transition is not a verdict on a presidency, and this page will not print one.

Event study: the conflict series with its transitions marked, and each transition shift against the placebo distribution
top: the monthly series with the transitions marked; bottom: the per-transition h=12 shift with its interval against the 5th and 95th percentiles of the placebo distribution
The rule the verdicts were read against

Per event, per outcome, at the h=12 focus horizon in the declared baseline variant: shift_detected when the per-event placebo p-value is <= 0.05 and the pre-trend placebo shift is smaller in magnitude than the post-window shift; not_interpretable_pre_trend when the pre-trend shift is at least as large; check_not_available when the pre-trend placebo cannot be run at all (it needs 120 months before the transition and the series is shorter behind the earliest event) — a missing check is reported as missing, not as a null; no_shift_detected otherwise. The family-wise p-value is published for every event regardless of its verdict, because four events read together are the largest of four whatever the rule says.

What this does not claim

Registry vintage 2026-11-01