Occupational exposures and risks of liver cancer among Shanghai female textile workers—a case–cohort study

Abstract

Background Liver cancer is the fifth most frequent malignancy worldwide. Viral hepatitis B and C, alcohol, and aflatoxin are the major established risk factors. Little is known about the aetiological contributions of occupational exposures, as previous occupational epidemiological studies of liver cancer suggest few agent-specific associations. We investigated associations of occupational exposures to dusts and chemicals in a cohort of female textile workers.

Methods Cancer incidence was determined among 267 400 female textile workers in Shanghai, China, who had been enrolled in an intervention trial of breast self-exam efficacy during 1989–98. Subjects were interviewed at baseline regarding basic demographics, smoking habits, alcohol consumption, and contraceptive practices. A case–cohort study of 360 liver cancer cases and 3186 age-stratified randomly chosen subcohort subjects was conducted within this cohort. Exposures to workplace dusts and chemicals were reconstructed from complete work history data, historical exposure monitoring data for selected agents, and a specially designed job-exposure matrix for the textile industry. Relative risks and dose–response trends were estimated by Cox proportional hazards modelling, adapted for the case–cohort design. Latency analyses with different lag years were also applied.

Results 2 095 904 person-years were contributed by this female cohort. The results of the case–cohort analysis revealed a protective effect of cotton fibre exposure years [adjusted hazards ratio (HR) = 0.64; 95% confidence interval (95% CI) 0.44–0.92] or endotoxin exposure (adjusted HR = 0.60; 95% CI 0.41–0.88) for the fourth quartile with significant trends for 20 year exposure lags.

Conclusions This study suggests that chronic exposure to endotoxin or some other component of cotton dust exposure may have reduced liver cancer risk in this population.

Liver cancer, occupational exposure, textile industry, job exposure matrix, cotton dust, endotoxin, latency analysis, female labourers

Background

Liver cancer was ranked fifth in incidence among all malignancies with an estimation of 564 000 new cases in the world in 2000.¹ It was also the fourth most common cause of cancer deaths worldwide with 400 000 deaths estimated by the International Agency of Research on Cancer (IARC) in 1999.² The highly variable geographic distribution of liver cancer and its pattern of occurrence in immigrant populations indicate that environmental factors are of major causative importance.³

Well-defined risk factors of liver cancer include hepatitis B virus (HBV), hepatitis C virus (HCV),^4–7 aflatoxins,^4,6,7 alcohol,^5,6 and oral contraceptives.⁸ Smoking, androgenic steroids, and diabetes mellitus are also suspected risk factors.^5,6

In developing countries in South-east Asia, the exceedingly high incidence (up to 100 cases per 100 000 person years) is associated with the carrier state of hepatitis B and probably a high concentration of the mould toxin aflatoxin B1 in food.⁹ In industrialized areas, alcoholic liver cirrhosis is the major recognized risk factor for liver cancer.^10–12 The crude incidence rates of liver cancer (ICD-9 code: 155) for women in Shanghai, China, were 14.8 and 15.3 per 100 000 during 1988–92 and 1993–97, reported by IARC.^13,14 Although specific characteristics of exposures and prevention policies for liver cancer in China have been described,¹⁵ further investigations are needed to identify remaining causes of this neoplasm.

The evidence for occupational risk factors, based primarily on studies of men, is fragmentary and inconclusive. Elevated mortality rates of liver cancer were detected among heavy construction equipment operators with the standardized mortality ratio (SMR) of 1.67 [95% confidence interval (95% CI) 1.06–2.50],¹⁶ chimney sweeps (SMR = 4.58; 95% CI 2.09–8.70),¹⁷ male chemical workers (SMR = 1.74; 95% CI 1.02–2.80),¹⁸ seamen (SMR for biliary passages and liver cancer, ICD-9 code 155 = 1.80; 95% CI 0.90–3.23),¹⁹ and painters (SMR = 1.25; 95% CI 1.03–1.50).²⁰ A meta-analysis by Chen and Seaton yielded an SMR of 1.20 (95% CI 1.04–1.37) for solvent exposures.²¹ In studies specific to the textile industry, two occupational cohorts in cellulose fibre production gave inconsistent results for biliary tract and liver cancers. One study found an SMR of 5.75 (95% CI 1.82–13.78) (both genders included),²² but the other reported an SMR of 0.81 (95% CI 0.02–4.49) for high exposure and 0.75 (95% CI 0.02–4.20) for low exposure to methylene chloride (for men only).²³ A census-linkage study from Shanghai, China indicated modest excesses among male textile machinery mechanics [Standardize Incidence Ratio (SIR) of 1.64 (P < 0.01) and bleachers, dyers, and textile product workers (SIR = 1.52; P < 0.01)].²⁴ Findings from several population-based case–control studies suggest some associations of occupational exposures with primary liver cancer. Male textile workers, not specified by job type, were found to have a higher risk of death from liver cancer [adjusted odds ratio (OR) = 3.08; 95% CI 1.22–7.76] in a case–control study from Texas, USA.²⁵ An international multicentre case–control study discovered an increased risk for hepatocellular carcinoma among female chemical industry workers (OR = 2.37; 95% CI 1.04–5.41).²⁶ We previously found slightly reduced liver cancer incidence among the cohort of women textile workers in Shanghai (SMR = 0.91; 95% CI 0.82–1.01),²⁷ although that analysis did not consider specific jobs or exposures.

Studies of female workers' occupational exposures and liver cancer have been scarce. Small sample size, inconsistent case definition, and incomplete confounding control precluded firm conclusions. Occupational classification has often relied on crude surrogate measures related to exposure based on the industry's characteristics, job or task assignment, and duration. This current study of female textile workers was designed to assess aetiological relations of liver cancer with specific exposures in a large well-characterized cohort.

Methods

Study design

The current case–cohort study was conducted within a cohort of female textile workers in Shanghai who were initially recruited into a randomized trial of breast self-examination (BSE) and followed-up for cancer incidence from 1989 to 1998. Human Subjects committees at the University of Washington, the Fred Hutchinson Cancer Research Center, and the Shanghai Textile Industry Bureau (STIB) reviewed and approved the study. Detailed descriptions of the established cohort are provided by Thomas et al.^28,29 In brief, the cohort was assembled during 1989–91 and included 267 400 women who were born between January 1, 1925 and December 31, 1958 (aged 30–64 at entry), employed in 519 STIB factories. At entry into the BSE trial, cohort members completed a baseline questionnaire that included data on demographic factors, smoking habits, alcohol consumption, and reproductive and gynaecological history.

Two tumour registry systems were used to identify liver cancer cases in this cohort during 1989–98: the Shanghai Cancer Registry, which was operated by the Shanghai Cancer Institute (SCI) and covered the metropolitan Shanghai area; and the STIB Tumour and Death Registry, which received three reports from each of its factories annually: all deaths due to cancer; all deaths due to other causes; and all women who had developed cancer within the past year. These reports were reviewed manually by registry staff to assure complete case identification. To assess completeness of case detection, data on women known to have liver cancer in the STIB were matched by computer to the SCI cancer registry data for concordance from 1989 to 1998. For records without a match in the SCI cancer registry, STIB staff were requested to review their medical records to confirm the diagnoses.²⁷ All women in the cohort were followed-up for vital status, working status (current or retired), work location within the STIB, and continued employment with the STIB.

Cases were defined as ICD-9 codes: 155.0 (primary malignant neoplasm of liver) and 155.2 (malignant neoplasm of liver, not specified as primary or secondary). The cases' diagnoses were verified by selected medical record reviews. A sample of 59 liver cancer cases was verified. Fifty-seven cases with available diagnosis data in medical records were diagnosed by imaging methods or microscope verification (96.6%). A subcohort group of 3194 women was randomly selected from the whole cohort, stratified by age. Sampling fractions for age groups were the component of weighting for non-case subcohort members in case–cohort analyses.

Exposure assessment

Exposures were assessed using qualitative and quantitative procedures. Two forms were developed and utilized by field workers and industrial hygienists in STIB. One was a ‘Work History Form’ that included information on individual employment period, work place, task description, and change in jobs. The other was a ‘Factory Profile Form’ for each factory used to record information on primary fibres utilized in the factory, manufacturing processes, the occurrence of hazardous exposures within the processes (chemical uses, biological and physical exposures) and quantitative exposure measurements when available. This information was collected for the entire period that the factory was in operation.

Exposure was assessed in two ways: with a job-exposure matrix (JEM) which classified various potential exposures dichotomously (yes/no) for each work process, and by quantitative estimation of exposure to cotton dust and endotoxin based on modelling current and historical exposure measurements. The JEM was based on a composite of two sources: (i) an a priori assessment of exposures by a process developed by Seattle-based Industrial Hygienists to classify fibre/process combinations according to the likely presence of various classes of dusts, chemicals, and physical agents (a priori classification), and (ii) the frequency of reporting of exposures by process on the Factory Profile Form (probabilistic classification). For probabilistic classification, agents that were reported with more than 30% frequency in a specific process were classified as ‘exposed’. The a priori and probabilistic classifications were then combined to create the ‘best evidence’ JEM with each process classified as exposed if either classification was positive.

In addition, quantitative estimates of exposures to cotton dust and bacterial endotoxin were made using quantitative data. A detailed description of the quantitative exposure assessment is available elsewhere.³⁰ Briefly, quantitative assessments of cotton dust exposure were estimated from historical measurements available from factory inspection reports. More than 2400 measurements were available, collected from 56 factories between 1975 and 1999. Along with factory and year, the variable major process and specific process were used to predict cotton dust exposure. An independent set of cotton dust measurements collected over the period 1981–96 in two of the factories included in this study were used to validate our exposure assessment model. The relative accuracy of the predictions was within 61% of the measurements.³¹ Endotoxin concentrations were assigned using the predicted cotton dust estimates and average concentrations of endotoxin per unit dust mass (EU/mg dust) in each major process from the studies by Christiani et al.^32–35

Exposure was then estimated for each study subject by combining the subject's work history with either the best evidence JEM or the quantitative estimates of cotton dust or endotoxin. Using the best evidence JEM, the number of years exposed to each fibre or other agent was estimated and classified as greater or less than 10 years of exposure. We estimated the hazard ratios (HRs) for duration of exposure to all agents other than cotton dust and endotoxin. Alternatively, we estimated HRs for cotton dust and endotoxin separately using quantitative cumulative exposure levels. The quantitative dust and endotoxin exposures were used to estimate cumulative exposure to each of these agents. Cumulative exposure was classified by quartiles among those individuals who were ever exposed. Potential confounding factors that were obtained from the baseline questionnaire included smoking, alcohol use, oral or injectable contraceptive use, and reproductive history.

Statistical analysis

Case–cohort analyses were conducted to estimate risks associated with workplace exposures,³⁶ adjusted for age and other potential confounders. Follow-up duration was defined as the period from the date of baseline questionnaire completion (starting to be at risk within the cohort) to the date of diagnosis (time to event) for liver cancer cases, date left STIB plants, date died from any cause, or the end of follow-up (December 31, 1998) for censoring subcohort members. The denominator weight in the pseudo-likelihood for subcohort members was the inverse of sampling fractions by age group (1/α_i) for HR estimations of case–cohort design suggested by Borgan.³⁷ Weights for the cases were set as one. Robust variance estimation was used.³⁸

Risks of liver cancer were estimated in relation to baseline characteristics of age, smoking habits, alcohol consumption, oral or injectable contraceptive use, and reproductive history. Cumulative exposure years of major fibre types (‘cotton’, ‘wool’, ‘silk’, ‘synthetic’, ‘other natural fibres’, ‘other mineral fibres’, and ‘non textile’) were calculated and treated as the major interest of predictive factors in analyses. Similar procedures were performed for other specific exposures, including dyes and pesticides. Exposure–response relations were estimated for cumulative exposures to cotton dust and endotoxin by the mid-point of each category.³⁹ We performed lagged exposure analyses, with lag intervals of 0, 5, 10, 15, and 20 years, to allow for disease latency.⁴⁰ Cumulative exposures were then categorized by quartiles of un-lagged exposures for the subcohort subjects. Women without the specific exposure experience were reference groups. The case–cohort analyses were performed by Stata/SE 8.0 for Windows, 2003.⁴¹

Results

There were 360 cases coded as ICD-9 code 155.0 and 155.2, eight of whom were identified in the chosen subcohort. The 3186 subcohort members without liver cancer served as the comparison group. The detailed results of univariate analyses for baseline characteristics are presented in Table 1. Age at baseline (in years) was a significant risk factor of liver cancer (HR = 1.10; 95% CI 1.09–1.11). None of the contraceptive practice or reproduction factors were associated with elevated liver cancer risk. Smoking was a rare habit among these women. We observed a moderately elevated risk among smokers, despite the low frequency of smoking in the cohort. Surprisingly, light alcohol consumption of less than once a month was found as a seemingly preventive factor for liver cancer, compared with the women who never drank (HR = 0.62; 95% CI 0.42–0.93). All the relative risk estimations in Table 1 were age-adjusted.

Table 1

Univariate analyses of risk factors in baseline questionnaire to the incidence of liver cancer

	Numbera
Risk factor	Case	Non-case subcohort	HR (95% CI)	P-value
Age in years at baselineb	–	–	1.10 (1.09–1.11)c	<0.01
Marital statusd
Married	325	2898	1.00 (–)	–
Divorced	1	21	0.43 (0.06–3.23)	0.41
Widowed	34	238	1.01 (0.69–1.49)	0.71
Contraceptive injection used
Never	350	3062	1.00 (–)	–
Ever	10	124	0.74 (0.43–1.25)	0.25
Oral contraceptive daily used
Never	322	2768	1.00 (–)	–
Ever	38	418	0.95 (0.66–1.37)	0.79
Number of pregnanciesd
0	17	110	1.00 (–)	–
1–2	75	777	0.87 (0.49–1.56)	0.65
>2	268	2299	0.71 (0.42–1.20)	0.20
Number of induced abortionsd
0	209	1642	1.00 (–)	–
1	93	977	0.86 (0.66–1.11)	0.24
>1	41	457	0.91 (0.64–1.31)	0.62
Number of spontaneous abortionsd
0	290	2610	1.00 (–)	–
1	37	365	0.83 (0.58–1.19)	0.30
>1	16	101	1.15 (0.67–1.99)	0.61
Regular smokingd
No (no >6 months)	333	3041	1.00 (–)	–
Yes	27	145	1.40 (0.91–2.16)	0.12
Total years of smokingd
None	333	3041	1.00 (–)	–
<5 years	4	20	1.45 (0.49–4.30)	0.50
5 years or more	23	125	1.40 (0.88–2.22)	0.16
Alcohol consumptiond,e
Never	321	2610	1.00 (–)	–
Less than once a month	29	432	0.62 (0.42–0.93)c	0.02
More or equal to once a week	10	144	0.56 (0.29–1.07)	0.08

a

Subjects with missing values were excluded from the analyses.

b

Treated as a continuous variable.

c

Statistical significance.

d

Adjusted for age at baseline.

e

P of test for trend <0.01.

There were no striking or consistent trends of risk for duration of exposure to any specific workplace agent (Table 2). Quantitative exposure–response trends for various lag intervals are shown for cotton dust and endotoxin in Tables 3 and 4, respectively. There are relatively weak inverse gradients for both exposures for lag intervals up to 15 years; however, there are inverse risk trends for cumulative exposures lagged 20 years. For the highest quartile of cumulative cotton dust exposure, lagged with 20 years, the relative risk estimate was 0.64 (95% CI 0.44–0.92). The corresponding result for endotoxin exposure was numerically similar (HR = 0.60; 95% CI 0.41–0.88).

Table 2

Relative risks for liver cancer by the estimation of best-evidence agent exposures by duration of exposure years

	Number
Exposure/years	Case	Non-case subcohort	HRa(95% CI)	P-value (trend)
Wool
Neverb	303	2743	1.00 (–)
<10 years	15	88	1.39 (0.78–2.47)
≥10 years	42	355	1.01 (0.71–1.42)	0.78
Other natural fibre except cotton, wool, and silk
Neverb	354	3118	1.00 (–)
<10 years	2	20	1.10 (0.25–4.81)
≥10 years	4	48	0.73 (0.26–2.05)	0.59
Any natural fibre
Neverb	104	814	1.00 (–)
<10 years	19	302	0.49 (0.29–0.81)c
≥10 years	237	2070	0.80 (0.62–1.02)	0.16
Synthetic fibre dust
Neverb	222	1937	1.00 (–)
<10 years	26	226	1.06 (0.68–1.64)
≥10 years	112	1023	0.95 (0.74–1.21)	0.69
Mineral dust, mineral fibre, and dust (NOC) and non-textile dust
Neverb	305	2676	1.00 (–)
<10 years	27	184	1.33 (0.87–2.03)
≥10 years	28	326	0.73 (0.48–1.10)	0.29
Solvents
Neverb	309	2738	1.00 (–)
<10 years	14	159	0.92 (0.52–1.63)
≥10 years	37	289	1.29 (0.89–1.86)	0.24
Bleaching agents
Neverb	357	3141	1.00 (–)
<10 years	1	15	0.98 (0.12–8.03)
≥10 years	2	30	0.59 (0.14–2.49)	0.48
Acids, bases, and caustics
Neverb	334	2928	1.00 (–)
<10 years	6	85	0.78 (0.33–1.83)
≥10 years	20	173	1.12 (0.69–1.81)	0.79
Dyes
Neverb	355	3092	1.00 (–)
<10 years	3	28	1.14 (0.34–3.85)
≥10 years	2	66	0.29 (0.07–1.18)	0.09
Inks
Neverb	355	3158	1.00 (–)
<10 years	3	6	5.30 (1.29–21.81)c
≥10 years	2	22	1.09 (0.25–4.71)	0.34
Resins monomers or coatings
Neverb	352	3102	1.00 (–)
<10 years	5	28	1.87 (0.71–4.94)
≥10 years	3	56	0.51 (0.16–1.65)	0.49
Metals
Neverb	334	2980	1.00 (–)
<10 years	7	71	1.08 (0.48–2.39)
≥10 years	19	135	1.40 (0.85–2.32)	0.20
Electromagnetic field or non-ionizing radiation
Neverb	126	1045	1.00 (–)
<10 years	21	323	0.56 (0.34–0.90)c
≥10 years	213	1818	0.90 (0.71–1.14)	0.51
Other exposures
Neverb	354	3115	1.00 (–)
<10 years	2	20	1.09 (0.25–4.79)
≥10 years	4	51	0.70 (0.25–1.96)	0.53
Silk
Neverb	340	3027	1.00 (–)
<10 years	7	33	2.00 (0.86–4.65)
≥10 years	13	126	0.93 (0.52–1.68)	0.86
Lubricants
Neverb	131	1238	1.00 (–)
<10 years	29	311	0.89 (0.58–1.37)
≥10 years	200	1637	1.05 (0.83–1.33)	0.65
Pesticides
Neverb	357	3141	1.00 (–)
<10 years	1	16	0.64 (0.08–4.95)
≥10 years	2	29	0.54 (0.12–2.32)	0.36

a

Adjusted for age at baseline, regular smoking, and alcohol consumption.

b

‘Never’ meant the exposure <1 year.

c

Statistical significance.

Table 3

Lag time analyses of the quantitative cotton dust exposure for the incidence of liver cancer

	Numbera
Cotton dust exposure (mg/m3 × years)c	Case	Non-case subcohort	HR (95% CI)	P-value (trend)
0-year lag
Reference (0)	119	1010	1.00 (–)
First quartile (0.31–55.85)	59	541	0.94 (0.68–1.32)
Second quartile (55.85–97.02)	54	541	0.83 (0.59–1.16)
Third quartile (97.02–143.50)	57	541	0.76 (0.55–1.07)
Fourth quartile (143.50–1069.33)	67	541	0.90 (0.65–1.24)	0.76
5-year lag
Reference (0)	120	1013	1.00 (–)
First quartile (0.31–55.85)	60	564	0.93 (0.67–1.29)
Second quartile (55.85–97.02)	52	542	0.79 (0.56–1.11)
Third quartile (97.02–143.50)	59	521	0.80 (0.57–1.12)
Fourth quartile (143.50–1069.33)	65	534	0.87 (0.63–1.20)	0.55
10-year lag
Reference (0)	121	1020	1.00 (–)
First quartile (0.31–55.85)	61	611	0.89 (0.64–1.23)
Second quartile (55.85–97.02)	50	517	0.77 (0.54–1.10)
Third quartile (97.02–143.50)	59	511	0.80 (0.57–1.12)
Fourth quartile (143.50–1069.33)	65	515	0.88 (0.64–1.22)	0.55
15-year lag
Reference (0)	125	1063	1.00 (–)
First quartile (0.31–55.85)	68	652	0.94 (0.68–1.28)
Second quartile (55.85–97.02)	47	481	0.75 (0.52–1.07)
Third quartile (97.02–143.50)	56	485	0.77 (0.55–1.08)
Fourth quartile (143.50–1069.33)	60	493	0.82 (0.59–1.14)	0.37
20-year lag
Reference (0)	135	1192	1.00 (–)
First quartile (0.31–55.85)	74	637	0.99 (0.73–1.34)
Second quartile (55.85–97.02)	59	454	0.92 (0.66–1.29)
Third quartile (97.02–143.50)	44	456	0.61 (0.42–0.88)b
Fourth quartile (143.50–1069.33)	44	435	0.64 (0.44–0.92)b	0.02b

a

Four cases and 12 subcohort members were not available for the analyses because of missing data.

b

Statistical significance.

c

Women who had never worked with cotton fibre were used as reference group. For the purpose of unified comparison bases, boundaries of exposure were categorized by the data without lag and maintained for the analyses on 5, 10, 15, and 20 years of lagging. Age at baseline, regular smoking, and alcohol consumption were adjusted.

Table 4

Lag time analyses of the quantitative endotoxin exposure for the incidence of liver cancer

	Numbera
Endotoxin exposure (EU/m3 × years)c	Case	Non-case subcohort	HR (95% CI)	P-value (trend)
0-year lag
Reference (0)	119	1010	1.00 (–)
First quartile (6.18–1505.22)	52	541	0.83 (0.59–1.18)
Second quartile (1505.22–2434.65)	65	541	0.95 (0.68–1.31)
Third quartile (2434.65–3550.67)	63	541	0.87 (0.63–1.21)
Fourth quartile (3550.67–145 360.10)	57	541	0.78 (0.56–1.09)	0.25
5-year lag
Reference (0)	120	1013	1.00 (–)
First quartile (6.18–1505.22)	54	574	0.83 (0.59–1.17)
Second quartile (1505.22–2434.65)	61	534	0.93 (0.67–1.28)
Third quartile (2434.65–3550.67)	59	522	0.86 (0.62–1.20)
Fourth quartile (3550.67–145 360.10)	55	531	0.77 (0.55–1.08)	0.28
10-year lag
Reference (0)	121	1020	1.00 (–)
First quartile (6.18–1505.22)	56	640	0.79 (0.56–1.11)
Second quartile (1505.22–2434.65)	63	513	0.93 (0.67–1.29)
Third quartile (2434.65–3550.67)	61	488	0.88 (0.63–1.23)
Fourth quartile (3550.67–145 360.10)	55	513	0.76 (0.54–1.07)	0.22
15-year lag
Reference (0)	125	1063	1.00 (–)
First quartile (6.18–1505.22)	65	678	0.87 (0.63–1.20)
Second quartile (1505.22–2434.65)	65	487	0.93 (0.70–1.33)
Third quartile (2434.65–3550.67)	50	469	0.72 (0.51–1.03)
Fourth quartile (3550.67–145 360.10)	51	477	0.72 (0.51–1.03)	0.23
20-year lag
Reference (0)	135	1192	1.00 (–)
First quartile (6.18–1505.22)	83	670	1.07 (0.80–1.44)
Second quartile (1505.22–2434.65)	59	478	0.83 (0.60–1.16)
Third quartile (2434.65–3550.67)	39	416	0.58 (0.40–0.86)b
Fourth quartile (3550.67–145 360.10)	40	418	0.60 (0.41–0.88)b	0.02b

a

Four cases and 12 subcohort members were not available for the analyses because of missing data.

b

Statistical significance.

c

Women who had never worked with cotton fibre were used as reference group. For the purpose of unified comparison bases, boundaries of exposure were categorized by the data without lag and maintained for the analyses on 5, 10, 15, and 20 years of lagging. Age at baseline, regular smoking, and alcohol consumption were adjusted.

Discussion

The main objective of this study was to identify potential workplace risk factors for liver cancer in the textile industry. In addition, data on non-occupational factors (smoking and alcohol use) were available and could be assessed as risk factors and controlled as potential confounders. Notable findings were observed for exposures to natural fibres, cotton dust, and endotoxin. We also found that contraceptive injections and oral contraceptives were not related to liver cancer incidence in this cohort. In Western populations, the association between oral contraceptive and liver cancer has consistently been demonstrated.^42–45 However, studies conducted among women in areas where HBV infection was endemic failed to reveal increased risk of liver cancer in users of oral contraceptives^46,47 or injectable contraceptives.⁴⁸ Our study findings were consistent with theirs on this issue. This could be due to a high prevalence of HBV infection among the population that contributed to a considerable portion of liver cancer incidence. Further studies focusing on non-HBV carriers in countries with high HBV prevalence level might be needed.

Our results revealed potentially protective effects from cotton dust and endotoxin, allowing for a 20 year latency. Other than cotton dust and endotoxin, significant results were also detected for ink and EMF with exposure duration of <10 years (Table 2), but these were unanticipated, and have no clear interpretations currently. Cotton textile workers represent a population with chronic exposures to endotoxins. Analyses of cotton dust by several researchers have shown high concentrations of bacterial endotoxins in cotton dust samples.^49,50 The observation of decreased cancer mortalities was systematically reviewed and the hypothesis that endotoxins in cotton has anti-carcinogenic effects has been suggested.⁵¹ Stimulation of tumour necrosis factor alpha (TNF-α) by endotoxin has been proposed as a contributing factor to anti-carcinogenesis since the 1970s.⁵² Specifically, endotoxins might operate through a CD₁₄-mediated immune response to enhance the effects of cytokines,^53,54 which have the effect of inhibiting metastasis of hepatoma cells in mice.⁵⁵ A synthetic endotoxin, DT-5461, was shown to inhibit liver and lung tumour metastasis in mice,⁵⁶ causing selective blood flow reduction to cancer tissue resulting in tumour necrosis in mice⁵⁷ and rabbits⁵⁸ through TNF production. Another synthetic endotoxin, ONO-4007, has been shown to contribute to enhance immunocompetence for antitumour immunity⁵⁹ and to trigger TNF-α in hepatocellular carcinoma tissue in rat in vivo in a dose–response manner.^60–62 These effects occur in liver cancer tissue specifically.^63,64 Recently, a phase I human subject clinical trial for ONO-4007 of 24 patients with variety of cancers was performed, but no liver cancer case had been included as study subjects yet.⁶⁵ Consequently, there is currently insufficient evidence to conclude that endotoxin is an effective preventive factor for cancer. To our knowledge, this study is the first epidemiological investigation demonstrating that endotoxin exposure from inhaled cotton dust is inversely related to liver cancer incidence. The effect was most pronounced when exposures in the recent 20 years were ignored, which suggests early-stage effects. Endotoxin is ubiquitous in the working environment of cotton textile. The inverse relative risk gradients associated with cotton dust and endotoxin could conceivably be artefacts of a ‘healthy worker survivor effect’, if respiratory byssanosis-related symptoms of acute lung function decrements, cough, and phlegm,^66,67 caused workers to avoid exposure. We have no data on respiratory symptoms to address this potential bias. However, further analyses to compare the mean number of jobs by cases and subcohort members (mean number of jobs: 1.79 for cases and 1.87 for subcohorts; P for independent t-test: 0.21), and attempts of Tables 3 and 4 restricted to the women with one job only within their entire careers revealed similar results (fourth quartile of quantitative cotton dust exposure HR = 0.55; 95% CI 0.30–0.99; P for trend: 0.07; and fourth quartile of quantitative endotoxin exposure HR = 0.33; 95% CI 0.16–0.68; P for trend <0.01) as seen for all study subjects. Thus, survivor effect caused by selection of byssanosis-related symptoms seems unlikely to account for the inverse risk trends we observed.

Strengths of our study were data on incidence rather than mortality as in most other occupational cohort studies of liver cancer, a large case series (N = 360), the availability of data on some potentially important confounders (smoking, alcohol use, contraceptive methods), and complete work history data for all study subjects.^{16–20,22,23} Our ability to quantify historical exposures to cotton dust and endotoxin was also an important strength. The major limitation of this study was our inability to control for potential confounding of HBV, aflatoxin, and HCV. HBV infection and aflatoxin B₁ are established as the major risk factors for liver cancer in China^68,69 and prevention strategies for Chinese populations have thus been suggested accordingly.⁷⁰ Although these factors are strongly predictive of liver cancer risk, there is no reason to assume that their prevalence was correlated with workplace exposures to cotton dust and endotoxin. As such, the likelihood of important confounding would seem low. We did not observe an elevated risk associated with alcohol consumption, as might be expected. This may be due to relatively low quantities of alcohol consumed among self-reported drinkers in the cohort. Alternatively, the available data for alcohol consumption are not extensive, and may suffer from inaccuracies, which could have resulted in some residual, uncontrolled confounding. The extent of possible confounding is likely to be small, however, given the low frequency of alcohol consumption in the cohort. Another concern was that only 25.4% of our cases were diagnosed histologically, which reduced the precision of case definition. Most prior occupational studies of liver cancer have relied on mortality data, which is more prone to diagnostic error than incidence data.^{16–20,22,23,25}

Our findings of inverse liver cancer risk trends with cumulative exposures to endotoxin, in conjunction with plausible mechanisms derived from animal and in vitro studies, provide evidence strongly supportive of the role endotoxin may have in reducing the risk of liver cancer. The most prominent associations were detected when endotoxin exposures during the preceding 20 years were discounted, suggesting early-stage anti-carcinogenesis. Further work on this cohort will explore in greater depth the temporal patterns of exposure and risk, including analyses of exposure within specific time windows, to characterize possible mechanisms. There is also the possibility that components of cotton dust exposure other than endotoxin may have anti-carcinogenic potential, but most available evidence points towards a role for endotoxin. Our findings were specific to a cohort of Chinese women with a low prevalence of alcohol consumption and, thus, are not necessarily broadly generalizable. The effects of endotoxin on liver cancer risk deserve investigation in other workplace settings.

This work was supported by grant R01CA80180 from the US National Cancer Institution and Royalty Research Fund, University of Washington, Seattle, WA, USA, awarded in 2002. The authors thank Drs Fan Liang Chen, Yong Wei Hu, Lei Dan Pan, and Guan Lin Zhao for their leadership in supporting the follow-up of this cohort; Ms Wen Wan Wang and the BSE workers for their efforts of field work and medical record review; and Shirley Zhang for programming support. We also appreciate the suggestions from Drs Ziding Feng, Wenjin Li, Tak Sun Ignatius Yu, and Jean Woo for preparing the manuscript.