Correction

Hair Dye and Cancer: Four Findings Usually Quoted as One

The claim

“Hair dye causes cancer. The WHO classified it.”

The short answer

Four different studies asked four different questions and headlines fuse them. IARC classes working as a hairdresser as Group 2A and personal use of hair colourants as Group 3, not classifiable. The straightener finding is not a dye finding.

Evidence status

Contested

The Cosmetic Record · Published 2026-08-06 · Last checked 2026-08-06 · 9 sources · Corrections: none

1. Four separate findings, and how headlines fuse them

“Hair dye causes cancer, the WHO classified it” is one sentence assembled out of four separate bodies of work, no one of which says it. In publication order:

1. The IARC evaluations. A Working Group that met in Lyon in February 2008 published two separate verdicts in Volume 99 of the IARC Monographs (2010): one on occupational exposure as a hairdresser or barber, one on personal use of hair colourants. The second is the one that applies to a box of dye at home.¹
2. Zhang and colleagues, BMJ, 2020. A prospective cohort (exposure recorded before the disease, then followed forward) of 117,200 women in the Nurses' Health Study, over 36 years.⁵
3. Eberle and colleagues, Int J Cancer, 2020. 46,709 women in the Sister Study, breast cancer as the outcome, with a difference between Black and white participants that outside statisticians immediately disputed.⁶
4. Chang and colleagues, JNCI, 2022. 33,947 women from the same cohort, uterine cancer as the outcome, and an association with chemical straighteners — explicitly not with dyes.⁸

The fusion happens in two moves. IARC's occupational verdict is quoted and its personal-use verdict dropped, so a classification about a career in a salon is attached to a consumer. Then the 2022 straightener result is reported under the phrase “hair products”, which by the next news cycle has become “hair dye”. Both moves are avoidable by reading the abstract.

2. IARC: working as a hairdresser is not the same exposure as dyeing your own hair

Volume 99 of the IARC Monographs states its conclusions in five sentences, and they are worth reading in full because the halves are usually separated:

6.1 Cancer in humans

There is limited evidence in humans for the carcinogenicity of occupational exposures as a hairdresser or barber.

There is inadequate evidence in humans for the carcinogenicity of personal use of hair colourants.

6.2 Cancer in experimental animals

There is limited evidence in experimental animals for the carcinogenicity of hair colourants.

6.3 Overall evaluation

Occupational exposures as a hairdresser or barber are probably carcinogenic to humans (Group 2A).

Personal use of hair colourants is not classifiable as to its carcinogenicity to humans (Group 3).

The evidence behind the occupational verdict is specific, and it is not about breast or ovarian cancer. It is bladder cancer in men. Volume 99's summary of the human data reports that cohort studies showed an increased risk of bladder cancer among male hairdressers but not female ones; that in a Scandinavian cohort identified in the 1970 census and followed twenty years there was a significant 50% increase in men and a non-significant 10% decrease in women; and that most of more than twenty case–control studies reported increased risks in the range of 1.3–1.7 in male hairdressers.³

The rest of that section is rarely quoted: a pooled analysis of eleven studies in six European countries found no significant increase in either sex; smoking could generally not be adjusted for; and on breast cancer, many studies including the largest case–control and cohort studies showed no increased risk from professional use of hair colourants. One sentence governs all of it — the Working Group noted that the evidence mainly concerned exposures occurring before the 1980s, and often much earlier.³

None of this is new. IARC reached both conclusions in Volume 57 in 1993, and Volume 99 re-examined the question and arrived at the same two.⁴ The classification that circulates as breaking news is thirty-three years old.

3. What “Group 2A” and “Group 3” mean, and what they explicitly do not mean

The Monographs are graded on the strength of the evidence, not the size of the effect. The Preamble printed at the front of Volume 99 puts the distinction first:

A cancer ‘hazard’ is an agent that is capable of causing cancer under some circumstances, while a cancer ‘risk’ is an estimate of the carcinogenic effects expected from exposure to a cancer hazard. The Monographs are an exercise in evaluating cancer hazards, despite the historical presence of the word ‘risks’ in the title.

Group 2A therefore does not mean “more dangerous than Group 2B”. The Preamble states that probably carcinogenic and possibly carcinogenic “have no quantitative significance and are used simply as descriptors of different levels of evidence of human carcinogenicity”. Limited evidence is a positive association for which a causal interpretation is credible “but chance, bias or confounding could not be ruled out with reasonable confidence”. Inadequate evidence means the studies are of insufficient quality, consistency or statistical power to permit a conclusion about the presence or the absence of a causal association.²

That last clause decides how Group 3 should be read, and IARC says so itself:

An evaluation in Group 3 is not a determination of non-carcinogenicity or overall safety. It often means that further research is needed, especially when exposures are widespread or the cancer data are consistent with differing interpretations.

Group 3 is not a clearance and Group 2A is not a dose. Anyone quoting either label without the sentence that defines it is quoting a colour, not a conclusion — the same reasoning as Hazard Is Not Risk: How to Read Any Ingredient Score.

4. The largest prospective cohort, and both halves of what it found

The largest study of personal permanent hair-dye use is a prospective cohort of 117,200 women in the Nurses' Health Study, free of cancer at baseline and followed for 36 years, reported by Zhang and colleagues in the BMJ in 2020.⁵ It has two halves and coverage generally carried one of them.

The first half: ever-users of permanent hair dyes had a hazard ratio of 0.98 (95% confidence interval 0.96 to 1.01) for solid cancers excluding non-melanoma skin cancers, and 1.00 (0.91 to 1.10) for haematopoietic cancers overall, against non-users. Cancer-related death came out at 0.96 (0.91 to 1.02). Bladder cancer, melanoma, brain, colorectal, kidney and lung cancer, and the hormone-receptor-positive breast cancers, all showed no increase for ever-use.

The second half, in the paper's own words: “Basal cell carcinoma risk was slightly increased for ever users (n=22 560; 1.05, 1.02 to 1.08). Cumulative dose was positively associated with risk of estrogen receptor negative breast cancer, progesterone receptor negative breast cancer, hormone receptor negative breast cancer, and ovarian cancer.” The conclusion prints both halves in one paragraph: no positive association for most cancers or for cancer-related mortality, and the increases they did see “warrant further investigation”.⁵

The dose bands behind that second half are worth printing, because they are not a straight line. These are the paper's own Table 3 figures, for women of any natural hair colour: hazard ratios against non-users, with 95% confidence intervals, at three levels of lifetime use.⁵

Read down the columns rather than across the rows and the shape is the point. Only the heaviest band separates from the null in the breast-cancer rows; the middle band sits below the lightest in three of the four; and the ovarian row peaks in the middle and falls again. A dose that mattered would climb. This climbs, dips and climbs — a shape compatible with a dose effect and also with noise.

Two design features bear on how much weight the second half can take. More than 96% of the women had European ancestry, which the paper states and calls a limit on the generalisability of its findings. And it tested many cancer sites and subtypes across five exposure scales — the situation in which a few associations cross significance even under the null. The authors take the point up rather than ignore it, arguing that their five scales were complementary and their adjustment conservative.

5. The Sister Study: a difference between groups that is real and disputed

Eberle and colleagues analysed 46,709 Sister Study participants, aged 35–74 at enrolment between 2003 and 2009, over a mean 8.3 years of follow-up during which 2,794 breast cancers were identified.⁶ Permanent dye use in the previous twelve months carried a hazard ratio of 1.45 (95% CI 1.10–1.90) in Black women and 1.07 (0.99–1.16) in white women, heterogeneity p = 0.04. Personal straightener use carried 1.18 (0.99–1.41), with risk rising with frequency (p for trend = 0.02).

The design constraint is printed in the paper's own limitations: “By design, all study participants have a family history of breast cancer, which may limit the generalizability of these findings.” Every participant was recruited because a sister had been diagnosed. The authors argue this does not bias the internal comparison — a defensible position — but it is not a general-population cohort and the paper does not claim to be one.

The dispute was immediate and on the record. The Science Media Centre's expert reaction, published the morning the paper appeared, carries three named statisticians.⁹ Dr Michael Jones of the Institute of Cancer Research noted that “only just over 200 of the breast cancer patients were African American, limiting the power to draw conclusions about this group”. Dr Stephen Burgess of the MRC Biostatistics Unit wrote that the confidence intervals “were generally wide, and a zero association was always close to the range of the most likely values”. Professor Paul Pharoah of Cambridge went furthest: “The researchers carried out many analyses and the results highlighted in the abstract have been rather cherry picked.” That is commentary, not evidence, and is labelled as such here — but dropping it would misrepresent how the finding was received.

A published comment on the paper exists in the same journal, by Alipour; we could not read it, so we do not characterise its argument.⁷

6. Straighteners are not dyes — the finding most coverage dropped

The 2022 paper behind the largest wave of “hair products cause cancer” headlines is about a different product category. Chang and colleagues followed 33,947 Sister Study participants who had a uterus at enrolment, over an average of 10.9 years, during which 378 uterine cancers were identified.⁸ Ever-use of straightening products in the previous twelve months carried a hazard ratio of 1.80 (95% CI 1.12 to 2.88); use more than four times in that year carried 2.55 (1.46 to 4.45), p for trend = .002.

And then the sentence most coverage left out, from the same abstract: “Use of other hair products, including dyes and permanents or body waves, was not associated with incident uterine cancer.” The questionnaire separately recorded permanent, semi-permanent and temporary dyes, bleach, highlights, straighteners and perms. Straighteners were the category that moved. The authors call their result “the first epidemiologic evidence of association between use of straightening products and uterine cancer” and say replication is warranted.

The two categories are separated for a mechanistic reason: the straightener papers point at formaldehyde and formaldehyde-releasing chemicals, the dye papers at aromatic amines. Eberle's discussion notes that older straighteners used sodium hydroxide and thioglycolic acid salts and that formaldehyde-containing keratin treatments entered global markets in the early 2000s — which is also why an older null result and a newer positive one need not be in conflict.⁶

7. What a hazard ratio is, and what an absolute risk would have to look like

A hazard ratio is a relative rate, not an absolute chance. “A 45% higher risk” means the rate in the exposed group was about 1.45 times the rate in the unexposed group over the same period. It says nothing about how large either rate was, and if the underlying rate is small, a large ratio is still a small number of people. A ratio quoted without its baseline cannot be acted on.

Only one of these four papers publishes the baseline. Chang and colleagues estimated that among women who never used straighteners in the twelve months before baseline, approximately 1.64% were predicted to develop uterine cancer by age 70. The estimated risk was 1.18 percentage points higher for ever-users (95% CI for the risk difference 0.15% to 2.54%) and 2.41 points higher for frequent users (0.52% to 4.80%). They give the numbers needed to harm as 85 and 42 — one additional uterine cancer for every 85 ever-users, or every 42 frequent users.⁸ That is what an absolute risk looks like, and it is conditional on the association being causal, which the paper does not claim to have established.

Neither Zhang nor Eberle publishes an equivalent figure. For hair dye, there is no absolute-risk estimate in these papers to quote, only hazard ratios — and a hazard ratio of 1.07 with an interval running from 0.99 to 1.16 is compatible with no effect at all.

8. What would change our reading

This page is marked contested rather than settled. Four things would move it.

1. A new IARC evaluation using post-2000 exposure data. Volume 99's Working Group met in 2008 and recorded that its human evidence mainly concerned exposures before the 1980s; the same volume notes that in 2007, 135 individual ingredients were no longer allowed in the European Union for use in hair dyes.³ A classification resting on pre-1980 salon exposure describes products that in large part no longer exist.
2. Replication of the straightener–uterine-cancer association in an independent cohort. One prospective finding, called the first by its own authors, is a hypothesis with a confidence interval attached.
3. An absolute-risk estimate for the dye findings, of the kind Chang produced for straighteners. Until one exists, the receptor-negative breast-cancer and ovarian signals cannot be weighed against anything.
4. A subgroup analysis in Black women with adequate statistical power. The Eberle difference rests on just over 200 cases — the criticism the Science Media Centre respondents made, and the gap the authors themselves flag: “Future studies that are powered to robustly evaluate the possibility of differences by race and tumor subtype are needed.”⁶

What would not change it: another headline built by quoting IARC's occupational verdict and omitting its personal-use verdict from the same page of the same document.