Personal use of hair dyes and risk of leukemia: a systematic literature review and meta-analysis

Literature searches

We followed standardized PRISMA protocol and performed a systematic literature review to estimate the pooled relative risk of leukemia among personal hair dye users. Studies were identified by electronic database searching of PubMed, Web of Science, and Embase from their respective inception dates, using the search terms “hair” and (“dye” or “color” or “colour”) and (“leukemia” or “leukaemia” or “myeloid” or “lymphocytic” or “myelogenous” or “lymphoblastic”). Findings were supplemented with references manually obtained from the search results. Furthermore, we systematically reviewed studies that cited the articles that met our inclusion criteria. The latest search was conducted in May 2017.

Study selection

Two independent individuals reviewed and screened the title and abstract results to identify potentially relevant articles (Fig. 1). Selection criteria for full article review included (1) data on personal hair dye use, (2) all types of leukemia or subtypes of leukemia studied, (3) cohort or case–control studies, (4) reported risk estimates and 95% confidence intervals (95% CI), or available data for the calculation of measures of association. We excluded occupational studies examining hairdressers, preleukemia health endpoints (i.e., essential thrombocythemia and myelodysplastic syndromes), perinatal exposure to hair dye use, reviews, editorials, and articles not written in English. Additionally, studies reporting pooled risk estimates for both leukemia and general lymphoma endpoints grouped together were excluded, but studies reporting risk estimates for specifically the chronic lymphocytic leukemia/small lymphocytic lymphoma (CLL/SLL) endpoint were included due to the similarity of diseases. For studies examining the same cohort, preference was given to the more comprehensive analysis with longer length of follow-up or greater number of cases. The Kappa statistic was used to assess inter-rater agreement, and any disagreements were resolved by consensus with all authors.

Data abstraction

Data on study design, study population, primary exposure, subexposure, health outcome, risk estimates, confidence intervals, country of study, time period of diagnoses, time period of hair dye use data, number of cases and controls, cohort size, participation rate, and adjusted covariates were collected for each article. Study-specific data were abstracted by one reviewer and assessed for quality control by an independent reviewer.

Data synthesis and analysis

Each study was critiqued and analyzed for validity of study design and completeness of information based on potential for confounding, participation rate, loss to follow-up, methodology of interviews or questionnaires, and selection protocol. Sensitivity analyses were performed to determine the effect of study exclusion.

Fixed-effect and random-effects (in the presence of heterogeneity [P < 0.05]) models were employed to estimate the pooled relative risk (meta-RR) of all types of leukemia among all hair dye type users with 95% CI, weighted by the inverse of the study variance. Stratified analyses were performed by leukemia group (lymphocytic and myeloid), leukemia subtype (AML [acute myeloid leukemia and acute nonlymphocytic leukemia], CLL [chronic lymphocytic leukemia and small lymphocytic lymphoma], and ALL [acute lymphocytic leukemia and lymphoblastic leukemia]), duration of hair dye use (≤10 years or ≥15 years), time period of hair dye use (pre-1980 and post-1980), dark hair dye (black, brown, brunette, or classified as dark in paper), light hair dye (blonde or classified as light in paper), use of permanent hair dye, and gender. In these analyses, gender-, type-, and time period-specific risk estimates were included in the models when studies reported separate measures of association by males and females, permanent and semi-permanent hair dyes, and decade of hair dye use. No hair dye use was used as the reference group, and a minimum of four studies were required per stratified analysis. Cochran's Q test was used to examine study heterogeneity. Funnel plots (log relative risk and standard error) and Egger's test were performed to assess the potential of publication bias. All statistical analyses were performed using Stata software, version 14.2.

Overview of findings

This analysis evaluated 25 studies published between 1985 and 2016 that reported leukemia-specific risk estimates among hair dye users. The majority of selected studies reported relative risks by type of hair dye, duration/time period of use, gender, or leukemia subtype. A total of 149 individual risk estimates were considered in this analysis, 17 of which indicated a statistically significant increased risk of leukemia. When data from all studies were combined, we found that ever use of hair dye was associated with a nonstatistically significant increased risk of leukemia when compared to no use of hair dye. Findings also suggest that certain exposure characteristics, such as hair dye type (i.e., dark/permanent dye), time period of use (i.e., pre-1980), duration of use (i.e., ≥15 years), and male gender, are associated with a statistically significant increased risk of leukemia. However, the observed findings may be due to the introduction of bias and confounding by smoking, study design limitations, and historical formulations of hair dye products.

Risk by any hair dye

An unstratified meta-analysis indicated that use of any type of hair dye was associated with a nonstatistically significant increased risk of all types of leukemia, meta-RR 1.09 (95% CI: 0.97–1.22) (Table 2). This finding is in agreement with Takkouche 17 that reported a pooled RR of 1.12 (95% CI: 0.94–1.34) for all leukemia types and any dye users. The current analysis has tighter confidence intervals and included 10 case–control and two cohort studies not considered in the analysis performed by Takkouche.

Smoking is a suggested risk factor for the development of AML, ALL, and CML, potentially due to the presence of benzene, radioactive components, and other carcinogens present in tobacco and tobacco smoke 35-38. The majority (15/20) of studies in this analysis did not directly control for smoking, and the studies that did control for smoking did not provide risk estimates by leukemia type. When restricted to studies that adjusted for smoking, the magnitude of the risk of leukemia decreased. This provides evidence of potential confounding by smoking if smoking histories differed among users and nonusers of hair dye. For example, an increased prevalence of smoking among hair dye users could result in an increased risk of leukemia, independent of hair dye use, and overestimate the true effect. Because so few studies controlled for smoking, we were unable to perform smoking-adjusted stratified analyses. This requires further research, as it remains unclear if adjustment for smoking would materially influence the observed outcomes in the stratified analyses.

Data from all studies combined suggest that hair dye use generally does not have a causal effect on the development of leukemia, potentially due to insufficient doses of chemicals present in hair dyes, or the ability of skin cells to detoxify aromatic amines and attenuate systemic exposure 39, 40. Conversely, this crude unstratified meta-analysis may have diluted certain patterns of use that are associated with an increased risk of developing leukemia. Additionally, a significant amount of heterogeneity was observed in the unstratified model, likely due to the presence of various exposure classifications and study designs. Therefore, this study performed stratified meta-analyses by various exposure characteristics to understand the potential drivers of risk.

Risk by study design

When stratified by study type, ever use of hair dye was associated with a statistically significant increased risk of leukemia in case–control studies, which was not observed in cohort studies (Table 2). Hair dye use exposure was generally measured at baseline in cohort studies, which could introduce bias from exposure misclassification if hair dye use behaviors changed during follow-up. Specifically, nondifferential misclassification of binarily measured exposures would result in a bias in the risk estimate toward the null value, which could explain the observed reduction in the pooled risk estimate among cohort studies. Conversely, case–control studies are susceptible to recall bias if the reporting of exposures is systematically dissimilar between the cases and controls. Among case–control studies, if cases were more likely to report hair dye use exposures, then the pooled risk estimate would be biased away from the null and overestimate the true effect. It is therefore worth noting that (1) only case–control studies reported statistically significant increased risk estimates, and (2) the majority of studies included in the overall and subsequent stratified analyses were case–control studies.

Risk by hair dye type

Individuals who used dark-colored hair dye products were at a statistically significant increased risk of leukemia, whereas those that used light-colored hair dye products were not (Table 2). Dark hair dyes contain a higher concentration of primary aromatic amine intermediates and couplers in order to increase the shade of color 14. Hence, there are more reactive chemical species in dark hair dyes, which could explain the color-based difference in leukemia risk. Also, users of permanent hair dye were at a statistically significant increased risk of all types of leukemia, which was not observed among semi-permanent or nonpermanent hair dye users. This result differs from Takkouche 17, potentially due to the increased robustness of the underlying data in the current analysis. Permanent hair dyes undergo an oxidative reaction to impart color to the cortex of the hair shaft, which can result in oxidative damage 4. Oxidative damage has been associated with DNA mutations and cell proliferation during cancer initiation and progression 41. In contrast, semi-permanent or nonpermanent hair dyes do not penetrate the hair cortex or undergo oxidative reactions, but rather utilize van der Waals forces to adhere to the hair cuticle 4, 15.

Risk by time period and duration of use

This study also examined time period of hair dye use as a driver of leukemia risk. During the 1970s, the cosmetic industry incorporated aromatic amine ingredients (including 2, 4-diaminoanisole, 2-amino-4-nitrophenol, and 2, 4-diaminotoluene) into hair dye formulations that were shown to produce mutagenic activity and dermally penetrate humans, resulting in regulatory action in 1978 to ban certain aromatic amine compounds in the European Union 6, 15, 42-45. Additionally, in 1980, the Food and Drug Administration (FDA) required the presence of a warning label on hair dyes including 4-methoxy-m-phenylenediamine and 4,-methyl-m-phenylenediamone sulfate, stating that the product “contains an ingredient that can penetrate your skin and has been determined to cause cancer in laboratory animals” 46.

The implication of these time-varying exposures is that hair dyes prior to reformulation are potentially more carcinogenic than more recent hair dye formulations after the removal of certain mutagenic ingredients. Specifically, this study found that individuals who used hair dye products prior to 1980 were at a statistically significant increased risk of leukemia, which was not observed in individuals that used hair dyes after 1980 (Table 2). All underlying studies were age-adjusted case–control designs, so these variables likely did not differentially influence the observed outcomes. Among studies that reported “ever” use of hair dye, we were unable to stratify the underlying risk estimates in a systematic way by pre- and post-1980 dye formulations. Therefore, the current observed increased risks among studies may be driven by historical hair dye constituents. However, it should also be noted that the lack of a statistically significant association for more recent post-1980 hair dye use may be due to insufficient induction and latency periods since first exposure, or reduced cumulative lifetime exposures.

This analysis also provided evidence that duration of use may be a factor in the risk of leukemia, as use of hair dye for ≥15 years was associated with an increased risk of leukemia, while use of hair dye for ≤10 years was associated with a decreased risk of leukemia (Table 2). All underlying studies adjusted for age. However, these findings could also be an artifact of prior formulations, as individuals using hair dyes for ≥15 years may have used hair dye products prior to 1980. Similar to the time period stratification analysis, lack of an increased risk for the ≤10 years group could be due to latency, which has been reported to range from 1.5 to 15 years for leukemia 47.

Risk by leukemia type and gender

Lymphocytic leukemia develops from lymphocytes while myeloid leukemia arises from primarily granulocytes or monocytes. Similar to Takkouche, we report that hair dye use was associated with an increased risk of lymphocytic leukemia (Table 2). This trend was also observed when further stratified by leukemia subtype, with hair dye use being associated with a higher magnitude of risk for ALL and CLL than AML. We were unable to perform a separate model for CML due to an insufficient number of studies. Research has suggested that lymphocyte cells may be susceptible to hair dye constituents, as it has been noted that DNA damage in lymphocytes was slightly higher in volunteers after hair dying, and that single-strand DNA breaks were produced in lymphocyte cells exposed to para-phenylenediamine 48, 49.

Findings also suggest that male users of hair dye are at an increased risk of leukemia, which is an area of research that requires further investigation. This may be due to different hair dye use exposure patterns (i.e., frequency of use, differences in dye types, etc.); however, as there were only four studies that provided male-specific risk estimates, we were unable to further stratify subanalyses by gender. It should be noted that all studies involving men were case–control studies.

Strengths and limitations

The main strength of this study is that it provides a pooled estimate of leukemia risk among hair dye users across various study populations. Additionally, the inclusion of more studies allowed for the further investigation of exposure characteristics, including hair dye type and duration/time period of use. A methodological limitation was less accurate exposure classification in studies that assessed the risk of leukemia due to a suite of past occupational and lifestyle exposures, and therefore only collected limited data on the history of lifetime hair dye use 20, 27, 28, 34, 50, 51. Specifically, studies that only collected information on “ever” use of hair dye were unable to adequately characterize exposure profiles that may influence the risk of disease, as shown in the above stratified analyses.

This analysis is also limited by the studied disease classifications, with some studies reporting risk estimates for all leukemia types combined rather than by specific leukemia subtypes. The most informative summary risk measures should be analyzed by leukemia subtype due to different etiologies. However, we were unable to perform meta-analyses for all leukemia subtypes, and were unable to further stratify subtype meta-analyses by hair dye type, gender, etc. This body of literature would benefit from additional studies by specific leukemia subtypes. Additionally, the underlying studies did not provide data that enabled us to perform a robust dose–response analysis. While we used duration of use as a proxy of cumulative dose to assess the dose–response relationship, future studies should attempt to characterize different dose levels.

The issue of dependency also arises during the analysis of underlying case–control studies. In studies that calculated risk estimates for non-Hodgkin lymphoma, multiple myeloma, and leukemia, the same control populations were used in all calculations despite differing disease outcomes. Therefore, these calculations in the underlying studies may not be considered independent, and multiple comparisons may result in observed statistically significant associations that do not exist 52.