Independent research, published in full.

Olaf C. Woodall, Nostia's co-founder and CEO, writes quantitative research separately from the company — sole-authored, unfunded, and not about the product. What it finds is reported as it came out, including the confidence intervals, the null results, and the pre-registered predictions that failed.

Studies

Newest first. Each is free to read, download and share under the licence it carries.

Research report — not peer reviewed

9 October 2026

Where presidential campaigns go, and what travels online: a pre-registered search of US county election returns, 2016–2024

Olaf C. Woodall

One argument about national campaigns holds that candidates can skip small places, because the internet carries their message there anyway, and should campaign in big cities instead. Its premises were tested on every US county outside Alaska and Connecticut in the 2016, 2020 and 2024 presidential elections. The same data were then searched, one pre-registered hypothesis at a time, for what decides where campaigns hold events and how county votes move.

Campaigns already go where people live, and a visit barely moves its own county's vote, so these data give no vote-based reason to prefer big cities over small states. Counties did swing with the distant places they are tied to on Facebook, but both swings happen in the same election, so whether online ties carried the change is left open.

What it found

One campaign event, its county's vote: +0.07 pointsDemocratic event · 95% CI −0.20 to +0.34 · Republican event −0.20 (−0.41 to +0.01) · half a point excluded
Population, per log unit: +0.158 events95% CI 0.112 to 0.204 · within contested states · same sign in 2016, 2020 and 2024
Events hosted last election, per event: +0.503 events95% CI 0.370 to 0.636 · beyond population, lean, closeness and the last swing

Data: MEDSL county returns, FairVote event trackers, Meta Social Connectedness Index, Census
Sample: 3,104 counties in 2016, 2020 and 2024: 9,312 county-elections and 863 public events
Outcome: Two-party swing, turnout change and public campaign events per county
Design: Within-state comparisons, design-based regression, 11 pre-registrations
Licence: CC BY 4.0

Nineteen hypotheses were registered for the final stage of the search, which aimed to find three causes. Two passed every causal criterion and seventeen did not; all nineteen are reported. The strongest online pattern, a county's swing tracking that of its out-of-state Facebook friends (r = 0.72), is reported as a correlation.

Read the reportPDF · 453 KB · 12 pages

Working paper — not peer reviewed

19 September 2026

Self-rated oral health, but not tooth loss, is associated with postpartum depressive symptoms: a nationally representative cross-sectional study with a pre-registered negative control

Olaf C. Woodall

Poor oral health and depression travel together, but nobody can say which comes first, and almost nothing has been measured after a birth. Teeth settle the order: a missing permanent tooth is the record of years of disease, so it cannot have been caused by a depression that began after a birth in the previous two years. If oral disease drives postpartum depressive symptoms, tooth loss should show it. It does not — while how a woman rates her own teeth and gums, asked the same day as the symptoms, shows it strongly.

The association is carried by current, self-rated oral state rather than by accumulated oral disease, so it is unlikely to be a causal effect of that disease. Dental treatment should not be offered to this population on a promise of better mental health.

What it found

Self-rated oral health, per step worse: +0.68 PHQ-9 points95% CI 0.27 to 1.08 · p = 0.001 · OR 1.98 for screening positive
Missing permanent teeth, per tooth: −0.02 PHQ-9 points95% CI −0.14 to 0.09 · p = 0.68 · no gradient across categories
Missing third molars (negative control): +0.11 PHQ-9 points95% CI −0.14 to 0.36 · p = 0.39 · null, as pre-registered

Data: NHANES 2007–March 2020, six cycles
Sample: 861 women 1–24 months postpartum; 742 with a dentition examination
Outcome: PHQ-9 total score and the ≥10 screening threshold
Design: Cross-sectional, design-based regression, pre-registered
Licence: CC BY 4.0

Four predictions were registered before the analysis ran. The first — that tooth loss would be associated with symptoms — failed, and is reported as a failure.

Read the paperPDF · 344 KB · 12 pages

How the work is done.

Four things that are true of the work listed above.

Registered before the analysis runs: Hypotheses, models, covariates and numbered predictions are written down and committed before anyone sees a result. A prediction that fails is published as a prediction that failed.
Public data, deterministic code: NHANES and the county election, census and network data are public, and the analysis runs as a version-controlled pipeline rather than by hand, so the same inputs give the same numbers.
Intervals, not headlines: Every effect estimate is printed with its confidence interval, and a null is reported as a null. A result too imprecise to mean anything is described that way.
Nobody is identifiable: Only aggregates are published. The NHANES study shows no figure resting on fewer than five people — the same suppression rule the Nostia-Orgs console applies to a campus — and the election study uses public county totals, with no person-level data at all.