Paracetamol use and risk of ovarian cancer: a meta-analysis

Abstract

Aim

Ovarian cancer remains the most fatal gynaecological malignancy. Several observational studies have examined paracetamol as a potential chemopreventive agent. The nonconclusive nature of the epidemiological evidence prompted us to conduct a detailed meta-analysis of the studies published on the subject in peer-reviewed literature.

Methods

A comprehensive search for articles published up to 2004 was performed, reviews of each study were conducted and data were abstracted. Prior to meta-analysis, the studies were evaluated for publication bias and heterogeneity. Pooled relative risk estimates (RR) and 95% confidence intervals (CIs) were calculated using the random and the fixed-effects models.

Results

Eight studies (four case–control and four cohort studies), published between 1998 and 2004, were included. We found no evidence of publication bias or heterogeneity among the studies. The analysis revealed an inverse association between paracetamol use and ovarian cancer risk. This association was marginally significant assuming a random-effects model (RR = 0.84, 95% CI 0.70, 1.00), but not statistically significant assuming a fixed-effects model (RR = 0.90, 95% CI 0.80, 1.01). When the analysis was stratified into subgroups according to study design, the association was inverse in both case–control and cohort studies, but only in the former was it statistically significant. The sensitivity analysis strengthened our confidence in the validity of this association. Furthermore, our results provided evidence for a dose effect; ‘regular use’ was associated with a statistically significant 30% reduction in the risk of developing ovarian cancer compared with non-use (RR = 0.70, 95% CI 0.51, 0.95).

Conclusions

Our meta-analysis supports a protective association between paracetamol use and ovarian cancer, and provides evidence for a dose effect. However, the question of paracetamol’s potential association with ovarian cancer deserves further verification, since proof of chemoprevention would represent a major public health advance.

Introduction

Ovarian cancer remains the most fatal gynaecological malignancy. Although relatively uncommon, afflicting ∼ 1 of 60 women [1], the high mortality rate makes this disease a major health concern. The high mortality rate arises from the lack of an effective screening approach [2] combined with the limited success of the current therapy for advanced disease [3]. Strategies that focus on prevention may therefore provide the most rational approach for meaningful reductions in deaths attributable to ovarian carcinoma.

Analgesics have been suggested as potential chemopreventive agents. However, a recent meta-analysis failed to find evidence for a beneficial role of nonsteroidal anti-inflammatory drugs (NSAIDs) in the chemoprevention of ovarian cancer [4].

Several epidemiological studies have also examined paracetamol as a potential chemopreventive agent. However, the findings from these studies are inconsistent. Some indicated risk reductions in ovarian cancer with consumption of paracetamol, while others found no association.

Because of the widespread use of paracetamol, an association with a decreased ovarian cancer risk may have important public health implications. The nonconclusive nature of the epidemiological evidence prompted us to conduct a detailed meta-analysis of the studies published on the subject in peer-reviewed literature.

Materials and methods

Retrieval of published studies

To identify the studies of interest we conducted a computerized literature search. Sources included Medline (1966 to December 2004) and Science Citation Index Expanded (1970 to December 2004) databases. Search terms included: ‘paracetamol’ or ‘acetaminophen’ combined with ‘ovarian neoplasm’ or ‘ovarian cancer’ or ‘ovarian malignancy’. The title and abstract of studies identified in the computerized search were scanned to exclude any that were clearly irrelevant. The full text of the remaining articles was read to determine whether it contained information on the topic of interest. The reference lists of articles with information on the topic were reviewed to identify citations to other studies of the same topic. Reference lists of review articles were also reviewed to check for completeness of the assembled list of relevant publications.

Inclusion and exclusion criteria

The studies considered in this meta-analysis were case–control or cohort studies that evaluated exposure to paracetamol and ovarian cancer risk.

Articles were excluded from the analyses for any one of the following reasons: (i) they did not include paracetamol use as a risk factor for ovarian cancer, (ii) they did not provide an explicit description of paracetamol exposure, or (iii) there were insufficient published data for determining an estimate of relative risk and confidence interval. In studies with multiple publications from the same population, only data from the most recent publication were included in the meta-analysis, with reference in the text to the older publication. Inclusion was not otherwise restricted by study size or language.

We did not assess the methodological quality of the primary studies, since quality scoring in meta-analysis of observational studies is controversial. Scores constructed in an ad hoc fashion may lack demonstrated validity, and results may not be associated with quality [5–7]. Instead, we performed subgroup and sensitivity analyses as it is widely recommended [7–9].

We included all articles irrespective of publication length; that is, we did not exclude articles published as short reports, even though critical appraisal of such publications is limited. Hence, no study was rejected because of methodological characteristics or any subjective quality criteria.

We included in this meta-analysis studies reporting different measures of relative risk (odds ratio, incidence rate ratio, standardized incidence ratio). In practice, the three measures of effect yield very similar estimates of relative risk, since ovarian cancer is a rare occurrence.

Data extraction

Two reviewers abstracted the data independently to a predefined form. The following data were collected from each study: (i) publication data, first author’s last name, year of publication, and country of the population studied; (ii) study design; (iii) number of subjects; (iv) relative risks (RR) and 95% confidence intervals (95% CI); (v) case definition for ovarian cancer; (vi) definition of paracetamol exposure; (vii) control for confounding factors by matching or adjustments. Differences in data extraction were resolved by consensus, referring back to the original article.

In studies where more than one estimate of effect (RR) was presented, we chose the ‘most adjusted’ estimate, i.e. the estimate controlled for the largest number of potential confounders.

Statistical analysis

Two techniques were used to estimate the pooled relative risk estimates: the Mantel–Haenszel method [10] assuming a fixed-effects model and the DerSimonian–Laird method [11] assuming a random-effects model. In the absence of heterogeneity, the fixed-effects and the random-effects model provide similar results. When heterogeneity is found, both models may be biased [12].

Publication bias was evaluated using the Begg and Mazumdar adjusted rank correlation test [13] and the Egger regression asymmetry test [14].

To evaluate whether the results of the studies were homogeneous, we used the Cochran’s Q-test [15]. We also calculated the quantity I²[16], which describes the percentage variation across studies that is due to heterogeneity rather than chance. I² was calculated as I² = 100% × (Q − d.f.)/Q, where Q is Cochran’s statistic and d.f. the degrees of freedom. Negative values of I² were put equal to zero, so that I² lies between 0% and 100%. A value of 0% indicates no observed heterogeneity, and larger values show increasing heterogeneity [16].

Data were stratified into subgroups on the basis of study design. This was done to examine consistency across varying study designs with different potential biases. Homogeneity was assessed overall and within this stratification.

To evaluate the stability of the results of this meta-analysis, we performed a one-way sensitivity analysis. In this analysis, we evaluated the influence of individual studies by estimating the average relative risk in the absence of each study [17].

To assess any association between dose of paracetamol and the risk of ovarian cancer, drug exposure was grouped as ‘regular’ and ‘irregular’. ‘Regular use’ was the highest frequency, and ‘irregular use’ was the lowest frequency of drug use, as reported in the individual studies.

To assess any association between duration of paracetamol use and the risk of ovarian cancer, we used the available data from studies, which dichotomized duration to less and more than 10 years.

All P-values are two-tailed. For all tests, a probability level less than 0.05 was considered statistically significant.

This work was performed according to the guidelines proposed by the Meta-analysis of Observational Studies in Epidemiology (MOOSE) group [7]. Stata 6 software was used for the statistical analyses (STATA Corp., College Station, TX, USA).

Results

Search results

The primary computerized literature search identified 32 records. Seventeen articles were excluded because they were either animal studies (n = 1) or in vitro cell line experiments (n = 3), review articles (n = 2), or irrelevant to the current study (n = 11). We retrieved 15 potentially relevant manuscripts for further review. The full text was read and the reference lists were checked carefully. Finally, we identified 12 studies examining the association between paracetamol use and ovarian cancer [18–29].

Of these, three studies were excluded from the meta-analysis due to the rule for multiple publications from the same population [18, 19] or because they evaluated the use of analgesics as a risk factor for ovarian cancer, without differentiating between paracetamol, aspirin and non-aspirin NSAIDs [20].

In one case, two publications of the same authors had used the same population-based pharmacoepidemiology prescription database, in order to study cancer incidence [23] and mortality [21], respectively, among persons receiving prescriptions for paracetamol. In an attempt to choose the most reliable report, we decided to include in the meta-analysis the study of cancer incidence [23], as it appeared to be less prone to ‘confounding by indication’ bias. This specific type of bias, which is often present in epidemiological studies of commonly used analgesics, may have played a major role in the findings of the study investigating the association between paracetamol use and cancer mortality [30].

The remaining eight studies were included in the meta-analysis [22–29]. Of these, four were case–control studies [25–28] and the remaining four were cohort studies [22–24, 29]. There were no randomized trials. The number of cases ranged from 483 to 780 in the case–control studies, and from 10 to 1573 in the cohort studies.

Seven out of eight studies evaluated consumption of paracetamol and ovarian cancer incidence [22–28]. One study evaluated consumption of paracetamol and ovarian cancer mortality [29].

All studies [22–29] were controlled for potential confounding factors (at least for age), by matching or adjustments.

All case–control studies [25–28] had used noncancer controls. Among them, one study [27] had used two control groups, one cancer and one noncancer. However, we included in the meta-analysis the relative risk estimates derived from the analysis that had implicated the noncancer controls.

Seven studies [23–29] reported negative association (RR < 1.0) and one study [22] reported no association (RR = 1.0) between paracetamol use and ovarian cancer. Only two studies reporting relative risks < 1.0 had confidence intervals that did not include unity [26, 28].

The majority of the studies were conducted in the USA [22, 24, 26–29], but two were carried out in the UK [25] and Europe [23].

The publication dates of the studies included in the meta-analysis ranged between 1998 and 2004. Study designs, along with the estimated relative risks and 95% CIs, are shown in Table 1.

Table 1.

Studies included in the meta-analysis

Study	Study location	Study design	All subjects	OC cases	RR	(95% CI)	Control for potential confounders1	Source of exposure data	Comments
Lacey et al., 2004 [22]2	USA	Cohort	31364	116	1.0	(0.56, 1.8)	1-6	Questionnaire	Incident cases
Friis et al., 2002 [23]	Denmark	Cohort	13482	10	0.8	(0.4, 1.4)	1	Database	Incident cases; excluded persons with a prescription for aspirin/NSAIDs
Fairfield et al., 2002 [24]	USA	Cohort	76821	333	0.95	(0.53, 1.72)	1, 3, 7–12	Questionnaire	Incident cases
Meier et al., 2002 [25]	UK	C-C	2360	483	0.9	(0.6, 1.5)	1, 7, 9, 12	Database	Incident cases; censored exposure to paracetamol 1 year before cancer diagnosis
Moysich et al., 2001 [26]	USA	C-C	1641	547	0.56	(0.34, 0.86)	1, 4, 8, 11, 13, 14	Questionnaire	Incident cases
Rosenberg et al., 2000 [27]	USA	C-C	3350	780	0.9	(0.5, 1.6)	1	Interview	Incident cases
Cramer et al., 1998 [28]	USA	C-C	1086	563	0.52	(0.31, 0.88)	1, 3, 8, 12, 15–17	Interview	Incident cases
Rodriguez et al. 1998 [29]	USA	Cohort	616189	1573	0.98	(0.85, 1.13)	1, 3–6, 8, 11, 18–21	Questionnaire	Ovarian cancer deaths

OC, Ovarian cancer; RR, relative risk; CI, confidence interval.

¹Age; 2, ethnicity; 3, oral contraceptive use; 4, family history of ovarian cancer; 5, menopausal status; 6, duration of oestrogen use; 7, body mass index; 8, parity; 9, smoking; 10, postmenopausal hormone use; 11, tubal ligation history; 12, aspirin/NSAID use; 13, age at first birth; 14, presence of irregular menses; 15, education; 16, religion; 17, menstrual pain, headache, arthritic pain; 18, race; 19, age at menarche; 20, age at menopause; 21, family history of breast cancer.

²Numbers in square brackets, reference citation.

Meta-analysis

The P-values for the Begg and Mazumdar test and Egger test were P = 0.90 and P = 0.15, respectively, both suggesting that an assumption of no publication bias is reasonable. Similarly, the Cochran’s Q-test had a P-value of 0.19 (Q = 9.93 on 7 degrees-of-freedom) and the corresponding quantity I² was 29%, both indicating that for the results of the studies an assumption of homogeneity is reasonable (Table 2).

Table 2.

Meta-analysis results

						Tests of homogeneity			Tests of publication bias
No. of studies	RR	(95% CI)	RR	(95% CI)	Q value (d.f.)	P-value	I2	Begg’s P-value	Egger’s P-value
All studies	8	0.90	(0.80, 1.01)	0.84	(0.70, 1.00)	9.93 (7)	0.19	29%	0.90	0.15
C-C studies	4	0.69	(0.54, 0.89)	0.69	(0.52, 0.93)	4.00 (3)	0.26	25%	0.73	0.88
Cohort studies	4	0.97	(0.85, 1.11)	0.97	(0.85, 1.11)	0.40 (3)	0.94	0%	0.09	0.42
‘Regular use’	8	0.73	(0.57, 0.92)	0.70	(0.51, 0.95)	10.46 (7)	0.16	33%	0.27	0.15
‘Irregular use’	8	1.00	(0.88, 1.13)	1.00	(0.88, 1.13)	7.20 (7)	0.41	3%	0.17	0.16
Duration >10 years	3	0.90	(0.73, 1.11)	0.64	(0.34, 1.23)	8.87 (2)	0.01	77%	0.30	0.01
Duration ≤ 10 years	3	0.90	(0.77, 1.05)	0.80	(0.58, 1.10)	3.36 (2)	0.19	40%	1.00	0.12

RR, Relative risk; CI, confidence interval; d.f., degrees of freedom.

The statistical analysis revealed an inverse association between paracetamol use and ovarian cancer risk. This association was marginally significant assuming a random-effects model (RR = 0.84, 95% CI 0.70, 1.00, n = 8), but not statistically significant assuming a fixed-effects model (RR = 0.90, 95% CI 0.80, 1.01, n = 8) (Table 2). However, the random-effects model is generally thought to be more appropriate, because it provides a more conservative estimate of the pooled effect size.

To examine consistency across varying study designs with different potential biases, we stratified data into subgroups on the basis of study design. The inverse association was statistically significant among the case–control studies (random-effects model, RR = 0.69, 95% CI 0.52, 0.93, n = 4), but not among the cohort studies (random-effects model, RR = 0.97, 95% CI 0.85, 1.11, n = 4) (Table 2). Figure 1 illustrates the relative risks and 95% CIs from the individual studies and the pooled results.

Figure 1.

Analysis of studies, denoted by first author and publication year, which examined ovarian cancer and its association with paracetamol use. The relative risk and 95% CI for each study are displayed on a logarithmic scale. Pooled estimates are from a random-effects model

In the sensitivity analysis, the overall homogeneity and effect size were calculated, removing one study at a time (Table 3). This analysis indicated that, when the only study that evaluated consumption of paracetamol in relation to ovarian cancer mortality [29] was excluded, and the analysis was restricted to the studies that evaluated consumption of paracetamol in relation to ovarian cancer incidence, the P-value for the Q-test increased from 0.19 to 0.43 and the calculated effect estimate was identical in both a fixed- and a random-effects model (RR = 0.76, 95% CI 0.62, 0.93, n = 7).

Table 3.

One-way sensitivity analysis

Study excluded	Fixed-effects model		Random-effects model		Test of homogeneity P-value
RR	(95% CI)	RR	(95% CI)	P-value
None	0.90	(0.80, 1.01)	0.84	(0.70, 1.00)	0.19
Lacey et al. [22]	0.90	(0.80, 1.01)	0.81	(0.66, 1.00)	0.13
Friis et al. [23]	0.91	(0.80, 1.02)	0.83	(0.68, 1.02)	0.13
Fairfield et al. [24]	0.90	(0.80, 1.01)	0.82	(0.66, 1.01)	0.13
Meier et al. [25]	0.90	(0.80, 1.02)	0.82	(0.66, 1.01)	0.13
Moysich et al. [26]	0.93	(0.83, 1.05)	0.93	(0.83, 1.05)	0.47
Rosenberg et al. [27]	0.90	(0.80, 1.02)	0.82	(0.67, 1.01)	0.13
Cramer et al. [28]	0.93	(0.82, 1.05)	0.93	(0.82, 1.05)	0.49
Rodriguez et al. [29]	0.76	(0.62, 0.93)	0.76	(0.62, 0.93)	0.43

To analyse any association between dose of paracetamol and the risk of ovarian cancer, drug exposure was grouped as ‘Regular’ and ‘Irregular’. ‘Regular use’ was the highest frequency, and ‘Irregular’ was the lowest frequency of drug use, as reported in the individual studies (Table 4). All studies [22–29] contributed to this analysis. There was evidence for a dose effect; ‘Regular use’ was associated with a statistically significant 30% reduction in the risk of developing ovarian cancer compared with non-use (random-effects model, RR = 0.70, 95% CI 0.51, 0.95, n = 8) (Table 2). In contrast, ‘Irregular use’ did not cause any reduction in the risk of developing ovarian cancer (random-effects model, RR = 1.00, 95% CI 0.88, 1.13, n = 8) (Table 2). Thus, there was no difference in the risk of developing ovarian cancer between irregular users and non-users.

Table 4.

Definitions of paracetamol use in the studies included in the meta-analysis

Study	Definition of irregular use	Definition of regular use
Lacey et al. [22]	Once a week—once a day for at least 1 year	More than once a day for at least 1 year
Friis et al. [23]	1 prescription for paracetamol during the study follow-up time	≥10 prescriptions for paracetamol during the study follow-up time
Fairfield et al. [24]	1–4days per month	≥5days per month
Meier et al. [25]	1–9 prescriptions for paracetamol recorded in the computerized medical history	≥30 prescriptions for paracetamol recorded in the computerized medical history
Moysich et al. [26]	1–6 tablets per week for at least 6months	≥7 tablets per week for at least 6months
Rosenberg et al. [27]	≥1day per week for at least 6months	≥4days per week for at least 6months
Cramer et al. [28]	1–6 tablets per week for at least 6months	≥7 tablets per week for at least 6months
Rodriguez et al. [29]	Occasional use during the month before enrolment in the study cohort	≥30 tablets during the month before enrolment in the study cohort

To assess any association between duration of paracetamol use and the risk of ovarian cancer, we used the available data from studies, which dichotomized duration to less and more than 10 years. Only three studies [26, 28, 29] contributed to this analysis. The calculated pooled relative risks for short and prolonged duration of use were both compatible with 1.0 (Table 2), providing no evidence of an association between the duration of paracetamol use and risk of ovarian cancer. However, the number of studies contributing to this analysis was very small (n = 3) and there was high evidence of heterogeneity among these three studies (Cochran’s Q-test P-value = 0.01) (Table 2).

Discussion

To the best of our knowledge, this is the first meta-analysis of published studies that evaluates paracetamol for ovarian cancer prevention. Our pooled results of observational studies suggest a protective association between paracetamol use and ovarian cancer (random-effects model RR = 0.84, 95% CI 0.70, 1.00). When the analysis was restricted to the studies that evaluated consumption of paracetamol in relation to ovarian cancer incidence, the heterogeneity decreased and the calculated effect estimate was identical in both a fixed- and a random-effects model (RR = 0.76, 95% CI 0.62, 0.93). Furthermore, our results provide evidence for a dose effect; ‘Regular use’ was associated with a statistically significant 30% reduction in the risk of developing ovarian cancer compared with non-use (random-effects model RR = 0.70, 95% CI 0.51, 0.95). On the other hand, the fact that the effect estimate was much weaker in the cohort studies questions the true nature of the association.

These results extend prior observational studies by permitting synthesis of data and providing more stable estimates of effect. Examined singly, existing observational studies did not show a consistent benefit (although none showed a harmful association), and most lacked statistical power to analyse this association adequately.

When meta-analysis of observational data is performed, consideration of study bias is critical [7]. Existence of a bias in favour of publication of statistically significant results is well documented in the literature [31–33]. However, the likelihood of important selection or publication bias in our results is small. We did not exclude any article during the identification and selection process, and the Begg and Mazumdar test as well as the Egger’s test revealed no relation between the estimate of relative risk and study size. So, we are confident that important publication bias due to preferential publication of large studies with significant findings is unlikely to have occurred. Similarly, the tests of heterogeneity indicated very little variability between studies that cannot be explained by chance.

Nevertheless, several limitations should be considered in interpreting the results of this meta-analysis. First, our search was restricted to studies published in indexed journals. We did not search for unpublished studies or for original data. However, we did not impose any exclusion criteria with regard to language, type of publication or quality. Second, the included studies were different in terms of study design and definitions of drug exposure. We tried to explore sources of heterogeneity conducting subgroup and sensitivity analyses. However, the summary effect estimates are based on sparse and heterogeneous data, and therefore should be interpreted with caution.

Third, the sources of exposure data differ among the individual studies. Six studies [22, 24, 26–29] used personal interviews or self-administered questionnaires that are subject to recall bias. Two studies [23, 25] used automated databases that provide detailed information on dates of use and types of drugs used. This information is equally good for cases and controls irrespective of the event of interest, since it was recorded prospectively. However, studies that used prescription databases lacked information on over-the-counter use. Fourth, observational studies lack the experimental random allocation of the intervention necessary to optimally test exposure-outcome hypotheses. Thus, results may have been confounded by several factors, given that each one of the studies included in our meta-analysis controlled for somewhat different confounding factors (Table 1). Furthermore, the possibility of residual confounding of lifestyle factors should be considered, as it is possible that there may be systematic lifestyle differences between women who use paracetamol compared with those who use other painkillers.

Fifth, observational epidemiological studies of drug exposure often encounter a specific type of confounding, which is called ‘confounding by indication’. It occurs when the underlying condition, for which the drug is prescribed, rather than the drug itself, increases or decreases the risk of the outcome under study [34]. It has been shown that paracetamol is particularly prone to this bias [21, 30]. Confounding by indication could arise due to prescription of paracetamol to treat early symptoms related to ovarian cancer. It should, certainly, produce a positive association between the drug use and ovarian cancer. If such bias exists, it would imply that the reduction in ovarian cancer risk among paracetamol users, shown in our meta-analysis, is underestimated. In other words, existence of ‘confounding by indication’ should mask the protective effect of paracetamol.

Although the epidemiological data currently available suggest a protective association between paracetamol use and ovarian cancer risk, our knowledge of the mechanisms underlying this association is incomplete. It is improbable that paracetamol reduces risk via a prostaglandin inhibitor pathway, because of its limited anti-inflammatory and prostaglandin inhibitory properties [35].

At present, the possible paracetamol-induced reduction of ovarian cancer risk may be attributed to three specific mechanisms of action [36]: (i) induction of specific reproductive atrophy due to its sex-steroid resembling phenolic ring [37, 38]; (ii) reduction of glutathione pools due to its NAPQI metabolite, which may play an important role in sterilizing premalignant ovarian lesions, since they are shown to lack proper levels of glutathione [39]; and (iii) inhibition of ‘macrophage migration inhibitory factor’ activity, which is necessary for proper ovulation [40]. Clearly, laboratory investigations should to be conducted to define further the biological mechanism by which paracetamol may influence risk.

In conclusion, synthesis of existing studies provides suggestive evidence for a potential role of paracetamol in the primary prevention of ovarian cancer. Although the risks (i.e. liver and chronic renal failure) of long-term use of paracetamol may outweigh the potential benefits in preventing ovarian cancer in populations at low risk for ovarian cancer, a randomized trial of paracetamol might be appropriate in high-risk populations. Mathematical models might be applied in the identification of women at high risk for ovarian cancer [41, 42]. Nevertheless, the question of whether the epidemiological evidence provides a firm basis for randomized clinical trials needs to be examined carefully, especially when the evidence comes from sparse and heterogeneous data.

Until the validity of and mechanisms for an association between paracetamol and ovarian cancer protection are better defined, this association cannot yet be regarded as one which would prompt a public health recommendation.

No outside funding was provided for this analysis.

Conflict of interest: None declared.