Orbiter of Fluoride and Health. This page carries the dose-threshold decision that the nucleus summarises: where on the fluoride exposure axis a child-IQ decrement is established, and whether the evidence reaches the 0.7 mg/L US community-water-fluoridation dose.
(National Toxicology Program, 2024) unless marked otherwise.
The finding is stratified, and the stratum is the decision
NTP concluded, at moderate confidence (OHAT scale: High / Moderate / Low / Very Low), that «higher estimated fluoride exposures (e.g., … drinking water fluoride concentrations that exceed the … Guidelines … of 1.5 mg/L …) are consistently associated with lower IQ in children», the association holding «across different study populations, study locations, study quality/risk-of-bias determinations, study designs, exposure assessment measures, and types of exposure data». Below that level the finding does not hold: «Associations between lower total fluoride exposure [… lower than … 1.5 mg/L …] and children’s IQ remain unclear.»
Two decision-relevant consequences:
- Above ~1.5 mg/L water fluoride (or the total-exposure equivalent): harm arm is live at moderate confidence. Someone on a private well or a naturally high-fluoride community system testing >1.5 mg/L has an evidenced reason to mitigate.
- At 0.7 mg/L US fluoridation: insufficient evidence, not a
null. NTP «do[es] not address whether the sole exposure to fluoride added to drinking water … at
0.7 mg/L in the United States and Canada … is associated with a measurable effect on IQ»
[EXTRACTED (NTP - Fluoride Neurodevelopment Monograph 2024) chunk 01]. The addendum’s newer meta-analyses «provide additional evidence of a dose-response relationship. However, uncertainty remains in findings at the lower fluoride exposure range.»
1.5 mg/L is edge-of-evidence, not a demonstrated knee
Per The Underivable Optimum and the threshold discipline (a guidance threshold is first hypothesised as the boundary of the data, not a curve feature): NTP chose the WHO 1.5 mg/L water guideline as its «higher»-exposure referent because «no alternative safety guidelines for total fluoride exposure exist» and it «represents a useful total fluoride exposure equivalent metric» (National Toxicology Program, 2024). The number marks where the confident evidence sits (most low-risk studies sampled populations above it), not an estimated inflection where the dose-response bends. So «harm below 1.5 is unclear» is a statement about the evidence density, not a demonstrated safe threshold — and the within-study non-linearity is inconsistent (see Yu 2018 below).
The evidence base
- 72 IQ studies; 19 low risk-of-bias (high quality), of which 18 reported an inverse association
(National Toxicology Program, 2024);
46 of the 53 higher-RoB studies also inverse (National Toxicology Program, 2024).
The 19 low-RoB studies: 15 populations, 5 countries (China, Mexico, Canada, India, Iran),
7,000 children; fluoride measured in water (n=15), urine (n=15), serum (n=2). (National Toxicology Program, 2024)
- Companion meta-analysis (NTP’s own «DTT Meta-analysis, Taylor et al. 2024» — see the one-backing
caveat on Fluoride and Health): group-level 59 studies (n=20,932) SMD −0.45 (95% CI −0.57, −0.33);
individual-level «regression slopes meta-analysis of 13 studies (n = 4,475 children) … found a
significant decrease in IQ of 1.63 points (95% CI: −2.33, −0.93; p-value <0.001) per 1-mg/L increase
in urinary fluoride». Direction «remained consistent when group-level exposure was restricted to <4
mg/L, <2 mg/L, and <1.5 mg/L».
- This companion MA is now held as the peer-reviewed primary — Taylor et al. 2025 (JAMA Pediatrics 179(3):282-292, CC-BY open access) (Taylor et al., 2025). Same figures, now with the full published dose-response table and a low-RoB stratum. Same NIEHS/NTP Division of Translational Toxicology stream (its ref 12 is the NTP monograph; shared authors — Taylor, Rooney, Bucher), so it is the primary form of the same evidence, not an independent second line — see the section below and the one-backing note on Fluoride and Health. The published table SHARPENS the «direction consistent <1.5» summary above, which turns out to be exposure-matrix-dependent at low dose.
Taylor 2025: the published dose-response, and the low-dose exposure-matrix split
The peer-reviewed MA bounds and operationalizes what NTP stated qualitatively (type-F refinement). Two distinct pooled objects, in different units — keep them apart:
- Individual-level regression slope (IQ points): «a 1-mg/L increase in urinary fluoride was associated with a statistically significant decrease in IQ score of 1.63 points (95% CI, −2.33 to −0.93; P < .001)» (13 studies, n=4,475) (Taylor et al., 2025). Restricted to low-RoB studies (11 of the 13): «the decrease in IQ score was 1.14 points (95% CI, −1.68 to −0.61; P < .001)» with low heterogeneity (I²=23%) (Taylor et al., 2025).
- Group-level dose-response slope (change in SMD per mg/L, NOT IQ points): the Table-2 β values below. Do not read these as IQ points — they are standardized-mean-difference changes.
The decision-critical object is the low-dose subgroup — where the slope goes as exposure is restricted toward fluoridation levels. It splits by exposure matrix (all-studies β, change in SMD):
| Restriction | Water fluoride β (95% CI) | Urinary fluoride β (95% CI) |
|---|---|---|
| All data | −0.15 (−0.20, −0.11)*** | −0.15 (−0.23, −0.07)*** |
| <4 mg/L | −0.22 (−0.27, −0.17)*** | −0.20 (−0.31, −0.08)** |
| <2 mg/L | −0.18 (−0.40, 0.03) NS | −0.08 (−0.15, −0.005)* |
| <1.5 mg/L | 0.05 (−0.36, 0.45) NS | −0.08 (−0.15, −0.003)* |
(Taylor et al., 2025) (Table 2). Low-RoB-only: water <1.5 mg/L β −0.32 (−0.91, 0.26) NS (point estimate inverse, wide CI crossing null; n=879 from 3 studies); urinary <1.5 mg/L β −0.08 (−0.15, −0.002)* (holds; n=4,179 from 4 studies).
The load-bearing reading — the matrix decides the low-dose verdict:
- Water-fluoride dose-response goes null below ~2 mg/L — «associations remained inverse when exposed groups were restricted to less than 4 mg/L and less than 2 mg/L; however, the association was null at less than 1.5 mg/L» (Taylor et al., 2025). The <1.5 water point estimate is even slightly positive (0.05), and stays non-significant even among low-RoB studies.
- Urinary-fluoride dose-response stays inverse and significant down to <1.5 mg/L, in both all-studies and low-RoB strata.
This SHARPENS NTP’s «unclear below 1.5» and the earlier «direction consistent <1.5» summary: below 1.5 mg/L the answer is exposure-metric-dependent, and Taylor attributes the water-based null to thin data, not a safe threshold — «There were limited data and uncertainty in the dose-response association between fluoride exposure and children’s IQ when fluoride exposure was estimated by drinking water alone at concentrations less than 1.5 mg/L» (Taylor et al., 2025), and «The ability to detect a true effect is reduced at lower exposure levels when exposure contrasts are diminished» (Taylor et al., 2025). Per The Underivable Optimum: the water <1.5 null is an evidence-density / power artifact (n=879, 3 studies), not a demonstrated inflection or a J-curve benign-below arm — do not read it as a safe threshold.
Why the matrix matters (the Measurement Error in Dietary Assessment mechanism, applied): urinary «fluoride concentrations include all ingested fluoride and are considered a valid estimate of total fluoride exposure» (Taylor et al., 2025), whereas water concentration under-captures total dose (non-water sources add noise that «may also decrease the precision of the effect estimates at lower fluoride concentrations in water» (Taylor et al., 2025)). So the urinary <1.5 signal is the exposure-completer measure — but it carries its own error, the spot-sample caveat below.
The magnitude framing the authors themselves attach: «Although the estimated decreases in IQ found in the regression slopes meta-analysis may seem small (1.63 IQ points per 1-mg/L increase in urinary fluoride) … a 5-point decrease in a population’s IQ would nearly double the number of people classified as intellectually disabled» (Taylor et al., 2025) — the absolute-vs-population point: a small individual shift is population-relevant at the distribution tails.
Appraisal honesty (symmetric standards)
The MA drew published critique (JAMA correspondence; the low-dose subgroup, spot-urine measurement error, study quality). Note the limits under symmetric standards — neither adopt a critique nor dismiss the MA by authority:
- Spot-urine noise, partly mitigated: «When compared with 24-hour urine samples, spot samples are more prone to the influence of timing of exposure and can be affected by differences in dilution» (Taylor et al., 2025), though «correlations between urinary fluoride concentrations from 24-hour samples and spot samples adjusted for urinary dilution have been described» (Taylor et al., 2025). The urinary <1.5 signal — the one that persists — rests on the noisier low-dose exposure measure; reverse causation and residual confounding remain live and unadjudicated here.
- Publication bias is in the high-RoB studies, not the low-RoB ones: the overall pool showed evidence of publication bias (Egger P<.001), but the low-RoB subset had none (Egger P=.56); trim-and- fill left the inverse association intact. (Taylor et al., 2025)
- Same-stream producer/funder: NIEHS intramural funding, and «NIEHS did have a role in the review approval of the manuscript and the decision to submit the manuscript for publication» (Taylor et al., 2025) — reinforcing that NTP + this MA are one backing, not two. No conflicts declared.
- No US data: «to our knowledge no epidemiological studies addressing fluoride exposure and children’s IQ have been conducted in the United States» (Taylor et al., 2025); the near-fluoridation evidence is the North American maternal-urine cohorts (transportability table below). Taylor notes >2.9 million US residents on community/private water >1.5 mg/L — the stratum where the harm arm is live — but the finding does not directly transport to the 0.7 mg/L US fluoridation setting.
Veneri 2023: an independent team, a near-shared study base, and an opposite RoB reading
A third fluoride-IQ meta-analysis enters here — Veneri, Vinceti, Filippini et al. (CREAGEN / University of Modena), an SR + one-stage cubic-spline dose-response (Veneri et al., 2023). It is an independent team (no author shared with NTP/Taylor), independently funded, using different analytic instruments (ROBINS-E risk-of-bias tool; cubic-spline dose-response) — but it pools a near-fully-shared primary-study base (its 33 studies sit almost entirely inside NTP’s 72 / Taylor’s 59; the census of fluoride-IQ primary studies is near-closed). So it is not an evidence-independent second line — see the qualified-E verdict on Fluoride and Health. What it adds is robustness-to-analytic-choices, and — decisively — an opposite reading of the low-RoB stratum.
Veneri’s pooled dose-response, in IQ points (keep the metric straight — these are IQ points, unlike Taylor’s Table-2 β values which are change-in-SMD):
- Water fluoride: «Dose-response analysis showed a substantially linear IQ decrease for increasing water fluoride above 1 mg/L, with −3.05 (95% CI −4.06; −2.04) IQ points per 1 mg/L up to 2 mg/L, becoming steeper above such level.» (Veneri et al., 2023) — a signal that turns on around 1 mg/L, i.e. entering the fluoridation-relevant range.
- Urinary fluoride: «A weaker and substantially linear decrease of −2.15 (95% CI −4.48; 0.18) IQ points with increasing urinary fluoride emerged above 0.28 mg/L (approximately reflecting a water fluoride content of 0.7 mg/L).» (Veneri et al., 2023) — note the CI crosses zero: NS.
Cross-MA parameter table (the same-quantity check before any comparison)
| Parameter | Veneri 2023 (ROBINS-E; spline) | Taylor 2025 (OHAT; slope MA) | Same quantity? |
|---|---|---|---|
| Urinary DR, IQ pts / mg/L | −2.15 (−4.48, 0.18), NS | −1.63 (−2.33, −0.93), sig (13 studies) | PARTIAL — both IQ pts per mg/L urinary F, same direction, concordant magnitude (~2 vs 1.6); but Veneri = group-level spline, Taylor = pooled within-study individual slopes. Significance differs (estimator + study set), not direction |
| Low-RoB restricted pool | MD +1.11 (−0.67, 2.89), NS — assoc VANISHES (n=1, Feng 2022) | slope −1.14 (−1.68, −0.61), sig, I²=23% (11 of 13) | SAME question (survives quality restriction?), but “low-RoB” is INSTRUMENT-RELATIVE (ROBINS-E admits 3 low / OHAT-stream 19) and metrics differ (high-vs-low MD vs regression slope). Opposite answers |
| Water DR, low dose | spline declines above 1 mg/L, steeper >2 | subgroup β (SMD): null <1.5 (+0.05 NS); inverse <2, <4 | conceptually concordant (water signal concentrates >1-2 mg/L); NOT same metric (IQ pts vs SMD-change β) |
| Bashash 2017 (ELEMENT) RoB | overall HIGH (confounding = High) | low — an NTP/Taylor low-RoB flagship | SAME study, OPPOSITE RoB verdict — the mechanical driver of the divergence |
The RoB-instrument friction (now filed as a standalone tension)
Elevated to -> Does the Fluoride IQ Harm Survive a Low Risk-of-Bias Restriction (this section is the nucleus’s summary of it).
Veneri’s RoB subgroup runs a monotone gradient the opposite way from the NTP/Taylor reassurance: «a MD of 1.11 (95% CI −0.67; 2.89) emerged for the only low RoB study, −4.27 (95% CI −6.44; −2.11) for moderate RoB studies, and −6.31 (95% CI −9.56; −3.06) for high RoB studies» (Veneri et al., 2023). Its cohort-only pool is also near-null: «this considerably milder association with IQ score decrease found in these high-quality longitudinal studies (MD −0.74) compared with the cross-sectional studies (MD −5.21) raises additional concerns about the potential influence of biases in the latter estimates» (Veneri et al., 2023). Veneri’s own summary: «a general trend towards weaker or null associations in the most carefully conducted studies» (Veneri et al., 2023).
Taylor’s low-RoB stratum does the reverse — the association tightens (I²=23%) and stays significant (slope −1.14) when restricted to its 11 low-RoB studies, and NTP’s headline is that the finding holds «across … study quality/risk-of-bias determinations». Both make a low-RoB-robustness claim, and they oppose — on a near-identical pool of data.
The friction does not trace to different data (the studies are shared); it traces to the RoB instrument deciding which studies count as high-quality. ROBINS-E, applied by Veneri, downgrades hard on unadjusted confounding, so it admits only 3 low-RoB studies and rates the ELEMENT/Bashash and MIREC/Till flagships HIGH and Moderate — exactly the near-fluoridation cohorts NTP/Taylor rate low and lean on. The Bashash rating flip is the smoking gun: the same study is a low-RoB anchor for one MA and a high-RoB study for the other. So the reassurance that «it holds in the high-quality studies» — the claim propping up confidence in the >1.5 mg/L harm arm — is instrument-dependent, not a property of the evidence. Under a stricter confounding bar the high-quality pool goes null.
This lowers, it does not resolve. Which RoB treatment is better calibrated for this literature —
ROBINS-E’s aggressive confounding downgrade vs OHAT’s — is itself unadjudicated here (a G-gap: it
would need reading each tool’s confounding criterion against these studies end to end). And Veneri
carries the symmetric caveat in the other direction: its low-RoB verdict rests on a single study
(Feng 2022), and «a major role of residual confounding could not be ruled out»
(Veneri et al., 2023). Net: the low-RoB-robustness
argument is contested from both sides, so confidence on the whole dose-threshold decision stays low.
Veneri’s internal tension: the spline vs its own quality strata
Veneri also holds a tension within itself, which cuts toward concern at low dose: its cubic-spline reads a decline entering the fluoridation-relevant range — the urinary spline turns down at ~0.28 mg/L (~0.7 mg/L water-equivalent), and «the results of the present meta-analysis seem to indicate that such an ADI [EFSA 1 mg/day] may not be deemed safe from a cognitive development perspective» (Veneri et al., 2023) — WHILE its RoB- and design-stratified analyses null the association. The authors’ own reading of those strata leans toward attenuation, not concern: they highlight that of the 4 cohort-designed studies, the 3 that «also adjusted for major confounding (Broadbent et al., 2015; Farmus et al., 2021; Till et al., 2020) found only a mild effect on children’s IQ» (Veneri et al., 2023). So Veneri does not cleanly corroborate either pole: its spline sharpens the low-dose concern the same direction as Taylor’s urinary <1.5 signal, and its quality-stratification undercuts the whole association. Read as a whole, it deepens the uncertainty rather than settling it.
The transportability problem — where the evidence was collected
Most low-RoB studies come from high-fluoride endemic regions (China, India, Mexico) at exposures far above US fluoridation. The support-factor question (does the effect transport to a 0.7 mg/L setting?) hinges on the few cohorts sampled near fluoridation levels — and those are the North-American pregnancy cohorts, the transportability bridge:
| Study (low-RoB) | Setting / exposure | Key IQ estimate (with CI) (National Toxicology Program, 2024) |
|---|---|---|
| Bashash 2017 (Mexico, ELEMENT) | maternal urine mean 0.90 mg/L | −2.50 full-scale IQ per 0.5 mg/L maternal urinary F (95% CI −4.12, −0.59) |
| Green 2019 (Canada, MIREC) | maternal urine mean 0.51 mg/L | −4.49 full-scale IQ per 1 mg/L maternal urinary F in boys (95% CI −8.38, −0.60); girls null (2.40; −2.53, 7.33); combined null (−1.95; −5.19, 1.28) |
| Till 2020 (Canada, MIREC) | water 0.13-0.59 mg/L | formula-fed −4.40 full-scale IQ per 0.5 mg/L water F (95% CI −8.34, −0.46); breastfed full-scale null |
| Yu 2018 (China, Tianjin) | water 0.2-5.5 mg/L | −4.29 IQ per 0.5 mg/L water F only between 3.40-3.90 mg/L (95% CI −8.09, −0.48); no significant assoc 0.2-3.40 mg/L |
Two things this table does that the bare conclusion does not:
- The fluoridation-level signal exists but is fragile and conditional — the near-0.7 cohorts (Bashash, Green, Till) find effects, but they are prenatal / infant exposures via maternal urine, frequently sex-specific (Green: boys only), and inconsistent across the IQ subscale and feeding mode (Till: formula-fed but not breastfed). This is not the same claim as «0.7 mg/L in tap water lowers IQ».
- Yu 2018 is the clearest within-study non-linearity: an effect concentrated at 3.4-3.9 mg/L with no significant association below 3.4 — cutting against a smooth monotone gradient down to fluoridation levels, and consistent with the >1.5 mg/L bound.
Contra-signal at fluoridation level: Broadbent 2015 (New Zealand, Dunedin cohort) followed children exposed to community water fluoridation / fluoride toothpaste / tablets and «found no clear differences in IQ scores of the subjects at 38 years of age» (National Toxicology Program, 2024) — a prospective null at a fluoridation-relevant exposure, part of why the below-1.5 stratum stays unclear rather than harmful.
Measurement: better than dietary recall, but not clean
Fluoride escapes the worst of Measurement Error in Dietary Assessment because urinary fluoride is a biomarker, not a food-frequency questionnaire — a genuine advantage over most nutrition exposures. But two errors remain: (i) many studies assign group-level (area) water-fluoride, an ecological measure that misclassifies individual intake; and (ii) total fluoride exposure is not the water number (dental products, black tea), so even a well-measured water concentration under-captures dose. Both push a true low-dose gradient toward the null, which is one reason the below-1.5 arm reads as «unclear» rather than «no effect».
What this changes
The decision is not «is fluoride safe» but at what exposure, for which developmental window. A private-well household above 1.5 mg/L faces a live, moderate-confidence harm arm (mitigation — filtration or an alternative source — is on the table). A pregnant woman on optimally fluoridated municipal water (0.7 mg/L) sits in the insufficient-evidence zone: the prenatal cohorts give a reason for concern but not an established effect, and the caries-benefit pole (unaddressed by NTP — see Fluoride and Health) weighs on the other side. The magnitude that would matter — a 1-2 point IQ shift per mg/L urinary F — is small at the individual level but population-relevant, and its transmission to 0.7 mg/L tap water is exactly what remains unestablished.