Research & evidence
A study of vaccine introductions, withdrawals and changes in uptake across 40 countries, set alongside the wider evidence.
Research paper28 October 2026
01 / Results from the paper
The analysis compares recorded autism with each jurisdiction’s preceding trend, around 404 schedule changes in 40 countries.
Repeated comparisons are combined within countries, then country summaries are pooled.
The pooled estimate
The model translates population changes using the change in vaccination uptake. A ratio of 1 means no change on this scale.
The vertical line marks 1: no change.
The estimate is close to 1, the no-change value. The two intervals show uncertainty with and without a shared-bias allowance.
| Cohort | Uptake-scaled model ratio | 95% random-effects interval | Countries | Comparisons |
|---|---|---|---|---|
| −6 | 0.976 | 0.869–1.096 | 37 | 28,031 |
| −5 | 0.985 | 0.876–1.109 | 36 | 29,716 |
| −4 | 1.010 | 0.903–1.130 | 38 | 31,017 |
| −3 | 1.021 | 0.912–1.143 | 38 | 32,682 |
| −2 | 1.016 | 0.910–1.134 | 39 | 33,445 |
| −1 | 0.994 | 0.887–1.114 | 36 | 34,716 |
| 0 | 0.965 | 0.854–1.090 | 35 | 31,724 |
| +1 | 0.967 | 0.863–1.082 | 38 | 32,516 |
| +2 | 0.994 | 0.891–1.108 | 39 | 32,329 |
| +3 | 0.977 | 0.874–1.091 | 38 | 32,307 |
| +4 | 0.981 | 0.879–1.095 | 37 | 33,662 |
| +5 | 0.989 | 0.885–1.106 | 36 | 32,819 |
| +6 | 0.972 | 0.870–1.087 | 37 | 31,979 |
Complementary analyses examine control outcomes, vaccine-and-place contrasts, vaccine pairs and local coverage.
For ratios, 1 means no association; 1.25 means 25% higher on the stated scale. The cohort-slope row reports percent change per birth cohort, with zero meaning no change.
| Analysis | Scope | Estimate | 95% interval | Measure and interval type |
|---|---|---|---|---|
| Schedule changes: primary model | 40 countries; 39,753 comparisons | 0.973 | 0.836 to 1.131 | Uptake-scaled ratio; random-effects interval |
| Autism compared with control outcomes | 25 countries; 49,966 comparisons | 0.913 | 0.759 to 1.098 | Autism-minus-control model ratio; random-effects interval |
| Additional cohort-slope analysis | 39 countries; 35,365 comparisons | -1.73% | -8.05% to +5.02% | Percent change per cohort; placebo-calibrated interval |
| Shared mean across 23 contexts | 23 contexts | 0.937 | 0.599 to 1.458 | Shared mean ratio; generalized-profile credible interval |
| Shared mean across 17 contexts | 17 contexts; 8,538 cells | 0.987 | 0.638 to 1.519 | Shared mean ratio; credible interval |
| Vaccine-pair interactions | 27 vaccine-pair contrasts | 0.967 | 0.881 to 1.063 | Interaction multiplier; credible interval |
| Kindergarten MMR coverage | 14 states | 1.002 | 0.950 to 1.057 | Autism share of special education per 10 coverage points; credible interval |
| Early-childhood disability records | 6 states | 1.002 | 0.761 to 1.321 | Autism share of disability classifications per 10 coverage points; modified Hartung-Knapp interval |
Each row answers its stated question. A context is a vaccine or schedule contrast in a particular setting; some observations are shared across rows. The primary model’s shared-bias interval is shown above.
The matched analysis compares autism with control outcomes in the same reporting systems. Context models summarize vaccine-and-place effects; the 17- and 23-context models cover different sets of comparisons. The interaction multiplier measures departure from the product of separate effects. Local MMR rows measure autism’s share of disability classifications. These rows are reported separately rather than pooled together.
Random-effects intervals allow country results to vary. Credible intervals describe uncertainty under the Bayesian model. Placebo calibration uses fake event dates. Modified Hartung–Knapp intervals adjust uncertainty when combining a small number of states. The cohort-slope analysis describes differences across successive birth cohorts, rather than accumulation within individuals.
02 / Robustness
Alternative trends, shorter follow-up, measured-uptake samples and country omissions all give estimates close to 1.
| Model or sample | Ratio | 95% random-effects interval | Countries | Overlapping comparisons |
|---|---|---|---|---|
| Linear trend: primary model | 0.973 | 0.836 to 1.131 | 40 | 39,753 |
| Quadratic trend | 0.960 | 0.828 to 1.112 | 39 | 34,775 |
| Spline trend | 0.980 | 0.848 to 1.133 | 37 | 28,296 |
| End follow-up four cohorts after the change | 0.963 | 0.829 to 1.118 | 39 | 36,108 |
| Measured vaccination uptake | 0.934 | 0.717 to 1.217 | 14 | 8,355 |
| Modeled vaccination uptake | 0.967 | 0.807 to 1.159 | 37 | 15,055 |
| Mixed measured and modeled uptake | 0.963 | 0.779 to 1.190 | 16 | 16,343 |
| Large increases in uptake | 0.984 | 0.843 to 1.148 | 38 | 20,514 |
| Countries with subnational data | 0.993 | 0.818 to 1.205 | 18 | 39,532 |
| Exclude vaccine mandates | 0.970 | 0.834 to 1.129 | 40 | 37,386 |
Every country-omission interval includes 1. Each table row uses its own retained sample; countries can appear in more than one uptake group. These intervals exclude the separate shared-bias allowance.
The result is stable across these choices. The estimate uses vaccination uptake to translate population changes to an individual-risk scale. Allowing a bias shared across countries widens its interval to 0.70–1.35, which includes a 25% increase.
The risk conversion uses measured or modeled uptake and a risk-mixture assumption. Its dose-response validation check fails, so the converted ratio remains conditional on that model. A value below 1 does not establish protection.
The retained timing audit flags 42 contributing California comparisons whose event dates and exact membership remain unresolved. That qualification applies to the current reported estimate. Outcome records capture diagnoses or service classifications, with differences in ascertainment across systems.
Sensitivity analyses assess stability within the specified models. Placebo checks, exposure validation and source timing address additional questions. The supplement and technical companion report their results and outstanding issues.
Read the diagnostic checks →03 / How the study works
Follow the same birth cohorts from vaccination to later outcome records. Compare the affected cohorts with their expected path, then combine country summaries.
Align vaccination early in life with autism records collected later for those same birth cohorts.
Compare at the same age, or within comparable age bands where the source allows. Counts from health, school and service systems each measure a different form of recorded autism.
A rising line alone cannot identify an effect. The comparison asks whether affected cohorts depart from the path expected without the exposure change.
Fewer children vaccinated → more vaccinated
With the same effect per vaccinated child and outcomes compared at the same age, an abrupt increase in uptake predicts an abrupt difference between birth cohorts. An effect that develops within a person does not, by itself, turn that cohort step into a ramp.
The horizontal axis is birth cohort; outcomes are measured later, at comparable ages. Calendar-year totals can smooth a cohort step by mixing cohorts. Recording changes are examined separately in the diagnostic checks.
Combine repeated analyses within each country, then pool the country summaries.
The declared random-effects model allows countries to differ.
04 / The wider evidence
Follow individual children. Examine specific vaccines. Look at whole populations. Each approach tests a different part of the claim.
Large cohort studies compare autism diagnoses in vaccinated and unvaccinated children. WHO’s review finds that the strongest studies do not support a causal link.
The research reviewed by WHO ↗MMR, thimerosal and aluminium have each been investigated. The evidence does not support them as causes of autism.
WHO’s autism overview ↗Schedule-change studies compare affected birth cohorts with their expected path. The findings above add this population-level test to the wider evidence.
Back to the study results ↑05 / Common questions
Autistic characteristics may become apparent in early childhood, when children also receive routine vaccines. An event occurring first does not establish that it caused what followed. Studies need a comparison that tests that explanation. About autism and early development ↗
Research has examined MMR, thimerosal and aluminium. WHO’s review concludes that the available high-quality evidence indicates that vaccines, including those containing thimerosal or aluminium, do not cause autism. Read the WHO assessment ↗
These data count children identified by health, education or service systems. The manuscript distinguishes changes in identification and reporting from changes in underlying incidence; a rise in recorded diagnoses alone cannot identify its cause. The paper’s discussion of diagnosis ↗
Yes. Safety research identifies specific risks: for example, fever-related seizures occur more often after a first MMRV dose than after separate MMR and varicella injections at the same visit. An autism conclusion does not replace the evidence about other outcomes. CDC: MMRV safety studies ↗
The paper examines vaccine pairs as well as introductions and withdrawals. The average pairwise interaction is close to 1. The separate schedule-load meta-regression uses ten contexts with similar early schedules; it does not use the full cohort dataset and provides little precision about load dependence. Combination and schedule analyses →
06 / Read further
Read the argument in full, examine the methods, or explore the underlying results and source records.
Findings, methods and interpretation.
28 October 2026 edition.
Detailed models and supporting analyses.
28 October 2026 edition.
Explore estimates by setting and vaccine, and examine the methods and source data.
Open the technical companion →