There is conflicting evidence on the association of HMG-CoA reductase inhibitors (statins) with various cancers. While some animal models, epidemiological studies and randomized trials suggest that statins may be carcinogenic,1, 2, 3 new laboratory data show that they have antiproliferative, proapoptotic and antiinvasive properties,4 and recent observational studies report inverse statin-cancer associations.5, 6, 7 In contrast, neither the Cholesterol Treatment Trialists' (CTT) Collaboration8 nor a recent metaanalysis by Dale et al.9 found any evidence that statin therapy influences cancer risk. Nevertheless, follow-up in the included trials was limited (often < 5 years) and an adverse effect on breast cancer risk could not be excluded.8, 9 Thus the issue remains topical and controversial.10
Large numbers of people are being exposed to statins (Fig. 1), the use of which is likely to increase substantially as guidelines recommend lower primary prevention thresholds.11 It is therefore of enormous public health importance to clarify the relationship of statin use with cancer risk.10 Previously published metaanalyses have differed markedly in their inclusion criteria and methods.3, 8, 9, 12, 13 A 2001 metaanalysis included only 5 randomized trials,12 1 report was pravastatin-specific,3 1 was limited to breast cancer13 and only 14 CTT trials (79,751 participants in cancer specific analyses) were included in a 2005 pooled analysis.8 The recent metaanalysis by Dale et al.9 contained several randomized trials not included in previous reports (n = 26 trials; 86,936 participants). However, the pooled effect estimates for site-specific cancers had wide confidence intervals, indicating continuing uncertainty, and investigation of between-study heterogeneity (differential sub-group effects on cancer risk, for example, by statin type) focused on all-cancer as an outcome, which may mask important, biologically plausible site-specific subgroup effects.14, 15, 16 Further, observational studies were not included in Dale et al.9; differences or similarities in observational results compared with randomized trials may provide important information, since trial populations are usually highly selected with relatively short follow-up, and the appreciably larger observational studies will generate more precise effect estimates for site-specific cancers. The metaanalysis presented here includes 26 randomized trials (103,573 participants) and 12 observational studies (826,854 participants), and explores important potential sources of between-study heterogeneity for several site-specific as well as all-cancer outcomes.
Inclusion and exclusion criteria
We included randomized trials or observational studies that measured all-cancer or site-specific cancer incidence or fatality associated with statins. A measure of the strength of the association must have been stated in the form of risk or odds ratios, or could be calculated from the raw data presented in the article. Trials were included if they compared statins with placebo, and observational studies if they compared statins vs. no statins. Observational studies that only compared statin use with other lipid lowering agents were excluded. We excluded studies with only highly specific populations, such as renal transplant patients and those with familial hypercholesterolaemia. Abstracts were not included unless later published as full papers to ensure that they had been peer reviewed.
We searched MEDLINE, EMBASE, Web of Science, ISI Proceedings and BIOSIS Previews bibliographic databases from inception to November 2005, using search terms for statins and cancers (Appendix), and screened electronic clinical trials registers and reference lists of included studies. The titles and abstracts of papers retrieved by the search were systematically reviewed against the inclusion criteria and the full manuscripts of all papers subsequently considered to be potentially eligible were obtained and read. Data were extracted by DRLB onto a specifically developed analytical database and were checked for accuracy by RMM. We identified duplicate publications by reviewing study name, authors, location, study population, dates and study design.
Risk ratios (odds ratios for case-control studies) for all-cancers and site-specific cancers amongst those exposed to statins compared to placebo (for randomised trials) or no statin treatment (observational studies) were computed, with their respective variance. Analyses of randomized trials and observational studies were conducted separately. The specific outcomes were incident and fatal all-cancers and the following site-specific cancers: breast, prostate, colorectum, lung, genito-urinary, melanoma and gastric. Fixed-effect models are reported throughout, as these reflect only the random error within each study and are less affected by small study bias,17 but we repeated all analyses using random effects models18 to determine whether the method of pooling the data influenced the results. We calculated the I2 statistic as a quantitative measure of the degree of inconsistency across studies.19 For studies reporting on different periods of follow-up, we included the results for the longest follow-up. Two observational reports presented cancer data for the colon and rectum separately5, 20; we computed a risk ratio for the predefined endpoint colorectal cancer by pooling within-study effect estimates for the 2 sites using fixed-effect models. We did not assess the quality of the included studies as quality assessment in metaanalysis is controversial and results can be highly misleading.21 Instead, we assessed small study bias by inspection of funnel plots and computation of Egger and Begg tests,17, 22 and we investigated the influence of potential confounding factors on associations reported in observational studies by comparing metaanalyses based on basic vs. maximally adjusted risk ratios where both these effect estimates were reported. Methodological aspects of individual studies were also discussed where heterogeneity was observed.
In a priori defined subgroup analyses,18 we repeated all analyses stratified by the following exposure variables, testing for difference between the subgroups using metaregression: (i) length of follow-up (≤5 years or >5 years, to allow sufficient time for the development of long latency cancers23); (ii) statin type, whether lipophilic (lovastatin, simvastatin, fluvastatin and atorvastatin) or other (pravastatin), since the increased lipid solubility of lipophilic statins may confer greater cell membrane permeability to influence cell proliferation, survival and motility24, 25 and (iii) statin potency: low (fluvastatin and lovastatin), medium (pravastatin) and high (simvastatin and atorvastatin).26 All metaanalyses and metaregression27 were conducted using Stata 9 (Stata Corp., College Station, TX).
Description of studies
The electronic search retrieved 3062 papers of which 60 were potentially eligible after review of their titles and abstracts (Fig. 2). A further 20 potentially eligible papers were identified after searches of reference lists and review of metaanalyses.8, 28 After the electronic search had taken place, 2 further eligible observational studies were published in 2006.29, 30 The full papers for the final total of 82 reports were obtained and, following review, 42 met the inclusion criteria; of these, 29 reports were based on randomized controlled trials and 13 were observational (6 case-control studies, 5 prospective cohorts, 1 historical cohort study and 1 nested case-control study) (Table I). Of the 42 papers, 3 reported data from the Airforce/Texas Coronary Atherosclerosis Prevention Study (AFCAPS/TEXCAPs).44, 45, 46 From this trial, we used data published in 199844 for incidence data, by Downs in 200145 for mortality data and by Clearfield in 2001 for breast cancer incidence.46 Two papers reported data from the Long-term Intervention with Pravastatin in Ischaemic Disease (LIPID) trial47, 48: we used data published in 1998 for breast cancer incidence and all cause mortality, and the 2002 publication for cancer incidence data (except breast). There were 2 reports from the UK General Practice Research database (GPRD): we used data published in 200260 for breast cancer incidence and the paper published in 200420 for the other cancer outcomes. Hence there were 38 individual studies, 26 trials involving 103,573 participants and 12 observational studies with 826,854 participants. Reasons for excluding 40 papers are in Figure 2.
Table I. Studies Included in the Metaanalysis (Ordered by Study Design and then Year Published)
First author (study name)
No. (cases/control) or [no. of cancers]
Mean follow-up (years)
Type of statin
Cancer outcomes measured
Age range (mean age); years
Trial Acronyms: ACAPS, Asymptomatic Carotid Artery Prevention Study; AFCAPS/TEXCAPs, Airforce/Texas Coronary Atherosclerosis Prevention Study; ALLHAT-LLT, The Antihypertensive and Lipid-Lowering Treatment to Prevent Heart Attack Trial; ASCOT-LLA, Anglo-Scandinavian Cardiac Outcomes Trial-Lipid Lowering Arm; CARDS, Collaborative Atorvastatin Diabetes Study; CCAIT, Canadian Coronary Atherosclerosis Intervention Trial; CCSS, Case-control Surveillance Study; CPS-II, Cancer Prevention Study-II Nutrition cohort; EXCEL, Expanded Clinical Evaluation of Lovastatin; GISSI, Gruppo Italiano per lo Studio della Sopravvivenza nell'Infarto Miocardico; GPRD, UK General Practice Research Database; HPS, MRC/BHF Heart Protection Study; KAPS, Kuopio Atherosclerosis Prevention Study; MAAS, Multicentre Anti-Atheroma Study; LIPS; Lescol Intervention Prevention Study; MARS, Monitored Atherosclerosis Regression Study; MECC, Molecular Epidemiology of Colorectal Cancer study; NHS, Nurses Health Study;SOF, Study of Osteoporotic Fractures; PLAC-IL, Pravastatin, Lipids, and Atherosclerosis in the Carotids; Post-CABG, Post Coronary Artery Bypass Graft; PROSPER, Pravastatin in elderly individuals at risk of vascular disease; PVAMC, Portland Veterans Affairs Medical Centre; REGRESS, Regression Growth Evaluation Statin Study; 4S, Scandinavian Simvastatin Survival Study; WHI, Women's Health Initiative Studies. NR, not reported.
Estimated that 53% of participants were using statins. NR = not reported.
We extracted data provided by the investigators of 7 early trials to the authors of one previous metaanalyses (from Table 8 of Ref.28), because cancer outcomes were not available as individual publications. These 7 trials were the Asymptomatic Carotid Artery Prevention Study (ACAPS),34 Kuopio Atherosclerosis Prevention Study (KAPS),36 Monitored Atherosclerosis Regression Study (MARS),32 Canadian Coronary Atherosclerosis Intervention Trial (CCAIT),33 Pravastatin Limitation of Atherosclerosis in the Coronary Arteries (PLAC-I),37 Pravastatin, Lipids and Atherosclerosis in the Carotids (PLAC-II)38 and the Multi-centre Anti-Atheroma Study (MAAS).35 In the Post-CABG trial, we used the results from the ‘moderate treatment’ arm rather than the ‘aggressive treatment’ arm.42 For Poynter et al.6 results from matched analyses were similar to the unmatched analysis and we used the data from the unmatched analysis. Eliassen et al.64 and Jacobs et al.29 provided additional data to us to allow inclusion of their studies in the metaanalysis.
Randomized controlled trials
There was no evidence of any association between statin therapy and all-cancer, breast, prostate, colorectal, lung, genito-urinary, melanoma or gastric cancer risk (Figs. 3a–3e and Table II). The heterogeneity in the estimate for the breast cancer association (I2 = 43%) was driven by the study of Sacks et al.41 and disappeared when this report was omitted (pooled estimate omitting Sacks et al. = 0.98; 95% CI: 0.76–1.26; I2 = 0%). The Sacks study was the first published trial to report on breast cancer outcomes in statin users and included only 13 cases (12 on statins).
The results for all outcomes were similar if analyzed separately for fatal cancer risk, apart from fatal colorectal cancers (Table II). In one study to report on fatal colorectal cancers, however, there were only 8 cancers,54 and the p-value implies that the association could easily have arisen by chance. Median length of follow-up in the trials was 3.6 years (interquartile range = 2.5–4.8; range = 1.0–10.4 years). Effect-estimates were similar (i.e. close to the null value) for all outcomes when the mean length of follow-up in the trials was 5 years or less vs. over 5 years (p-values for differences in effect-estimates by follow-up time ≥0.14) (Table III). Effect-estimates for all-cause and site-specific cancers were similar when metaanalyses were stratified by lipophilic agents vs. pravastatin or by statin potency (p-values for differences in effect-estimates by these subgroups ≥0.17).
Table II. Association of Statin Therapy with Various Cancer Outcomes in Published Randomised Controlled Trials
No. of studies
Risk ratio (95% CI)
Risk ratio calculations were not possible for 5 studies due to no deaths in either control (MARS, CCAIT, KAPS, REGRESS) or intervention (MAAS) groups.
Table III. Association of Statin Therapy with All-Cancer and Site-Specific Cancer Incidence by Subgroups (Duration of Follow-Up, Statin Type and Statin Potency) in Published Randomised Controlled Trials
Risk ratio (95% CI); Number of studies in pooled estimate
Obtained from a metaregression analysis to test the effect of the exposure subgroup on the pooled effect size18.
Median length of follow-up in the observational studies was 6.2 years (interquartile range = 4.6–6.8; range = 3.3–7.2 years). Metaanalyses of observational studies indicated some evidence of a small (8%; 95% CI: 1–15%) relative reduction in the risk of all-cancers associated with statin therapy (Fig. 4a and Table IV), although a random effects model was not statistically significant (random effects risk ratio = 0.92; 95% CI: 0.82–1.02; p = 0.10). There were insufficient studies to undertake a metaregression analysis of the statin-all-cancer association. There was no evidence of an association between statins and breast cancer risk, overall (Fig. 4b), or stratified by duration of follow-up or statin potency (p ≥ 0.7). There was weak evidence that lipophilic agents might be protective (risk ratio = 0.91; 95% CI: 0.80–1.03) compared with a possible adverse effect of pravastatin on breast cancer risk (risk ratio = 1.16; 0.93–1.43) (p for difference in effect-estimates = 0.08).
Metaanalysis of the risk of prostate cancer with statin therapy also indicated no evidence of any association (p = 0.38) (Fig. 4c and Table IV), although there was a high degree of inconsistency between studies (I2 value = 77%).19 A random effects model gave a similar result (random effects risk ratio = 0.93; 95% CI: 0.62–1.40). In metaregression analyses, there was no evidence that the statin–prostate cancer association varied by statin type or potency (metaregression p > 0.3), but there was weak evidence of small study bias (Egger p = 0.11). There were insufficient prostate cancer studies to undertake metaregression of the effect of duration of follow-up.
Table IV. Association of Statin Therapy with All-Cancer and Site Specific Cancer Incidence in Published Observational Studies
No. of studies
Risk ratio (95% CI)
I2 value (%)
For colorectal cancers, observational studies suggested a 14% relative reduction in risk in a fixed-effect model (Fig. 4d and Table IV), but no evidence of an association with statins in a random effects model (risk ratio = 0.84; 95% CI: 0.59–1.21; p = 0.36). Given the very high degree of inconsistency between study-estimates (I2 value = 89%), a random effects model might be more appropriate for this analysis. The statin–colorectal cancer association was driven by data from the Poynter et al. study, which showed a strong protective effect of statins in crude (odds ratio = 0.50; 0.40–0.63) and fully adjusted models (0.53; 0.38–0.74).6 This study was based in northern Israel with 69.2% of patients and 63.4% of controls being Ashkenazi Jews, known to be at increased risk of colorectal cancer and possibly affecting the generalizability of the results. There was no evidence of any statin–colorectal cancer association when the Poynter study was omitted from the metaanalysis (fixed-effect risk ratio = 1.02; 95% CI: 0.90–1.16; p = 0.7; I2 value = 54%). There was no evidence of small study bias in the statin–colorectal cancer association (Egger test, p = 0.8). There were insufficient studies to undertake a metaregression analysis of the statin–colorectal cancer association. Statin use was not associated with lung cancer (Fig. 4e and Table IV). Only 1–2 observational studies each reported on melanoma and gastric cancer outcomes, with no strong evidence for any association with statin use. One study also reported no evidence of any association with liver, female genital or haematological cancers.65 In studies that reported both crude and maximally adjusted effect estimates, controlling for a number of potential confounding factors made little or no difference to crude estimates (Table V).
Table V. Metaanalyses of Basic and Maximally Adjusted Risk Ratios in Observational Studies Reporting Both Outcomes
No. of studies
Risk ratio (95% CI)
Maximally adjusted model
Our metaanalysis of trial data found no evidence of any association between statin therapy and all-cause or site-specific cancer risk, including little evidence of between-trial heterogeneity in effect-estimates. We also showed that trial data were generally consistent with the results from observational studies. The observational studies did suggest the possibility of a small protective effect of statin therapy on all-cancer (8% relative reduction) and colorectal cancer (14% relative reduction) risks, although these findings were based on only 3 and 5 studies, respectively, there was no statistical evidence of any associations in random effects models (p values = 0.10 and 0.36, respectively), and the finding in relation to colorectal cancer was dependent on the inclusion of just one large study.6 Comparisons of observational vs. randomized clinical trial results for a number of other exposure-disease associations (e.g. β-carotene and heart disease risk) show that discrepancies in findings can occur because of uncontrolled (unmeasured) or residual confounding.66 Given the homogeneity of the null trial results for all cancers and colorectal cancers, it is possible that the small protective associations seen in the observational studies are the consequence of unmeasured confounding or other biases—problems known to plague even well-designed observational studies because exposures are not randomly assigned.66 For example, long-term statin users are healthier and more adherent to therapy and screening than nonusers,67 and may have been advised on a healthy diet, to increase physical activity and to stop smoking. Therefore, the slightly lower risk of all-cancers and colorectal cancer among statin users in observational studies may reflect the type of patients who seek and comply with statins, and other aspects of their medical management, rather than any statin-specific protective effect. The one large observational study to suggest an inverse association with colorectal cancers may not be generalizable.6
Our results are in line with a recent metaanalysis by Dale et al. of 26 randomized trials, in which statin use was not associated with all-cancer incidence or death, or with the following cancers: breast, prostate, gastrointestinal, colon, respiratory or melanoma.9 Dale's results were almost identical to ours, despite different inclusion criteria and statistical methods. Dale excluded trials with a study duration < 1 year and which enrolled < 100 patients,9 while we excluded trials in special populations, e.g. haemodialysis patients (in Dale's metaanalysis these were pooled with general population groups); second, we used fixed effect models and Dale employed random effects models.9 The novel aspects of our metaanalysis are that: (i) we included several large trials (mean trial size = 3240 participants) that were not in the Dale paper (EXCEL,31 ACAPS,34 Post-CABG42 and ALLIANCE58); (ii) we also analyzed 12 observational studies with 826,854 participants—these reports involved unrestricted populations prescribed statins in clinical practice, so may be more generalizable than randomized trials, and their typically larger sample size provided greater individual power to evaluate cancer-specific sites; and (iii) we did not limit our investigation of heterogeneity to the all-cancer outcome,9 but also undertook subgroup analyses for site-specific cancers.
This latter consideration is important as both Dale's and our study found a moderate degree of inconsistency among the trials reporting breast cancer outcomes, and it has been hypothesized that an increased breast cancer risk occurs with pravastatin (the only commercially available statin that is lipophobic25), but not lipophilic statins.14 This is because pravastatin might induce mevalonate synthesis in extrahepatic tissues, thus promoting breast cancer cell growth.68 Others argue that lipophilic statins have greater lipid solubility than pravastatin24 and more readily permeate cell membranes to exert potential anticancer effects.69, 70 Thus previously published metaanalyses may have diluted any association of statin use with site-specific cancer outcomes because trials of pravastatin and lipophilic agents were pooled.8, 9, 13 We found no evidence in our metaregression analyses of trials to support a differential effect of lipophilic agents compared with pravastatin for any site-specific cancer (p ≥ 0.17), although in observational studies there was weak evidence of an adverse effect of pravastatin on breast cancer risk (p for interaction by statin type = 0.08). [In Dale's paper, atorvastatin and fluvastatin were classified, along with pravastatin, as hydrophilic statins, but they are usually considered to be lipophilic,71 as in our classification; using the Dale classification made no difference to the results]. Our results for breast cancer are also in line with a metaanalysis published in December 2005 (after our search was completed) reporting no association (risk ratio = 1.03) of statin use with breast cancer based on 7 trials and 9 observational studies.13
The following limitations need to be considered in reaching conclusions based on current evidence. First, duration of follow-up was limited in both trials and observational studies; for example, only 3 of the 26 trials were followed up for over 5.0 years. Thus longer-term data are required to rule out long-latency effects. In metaregression analyses, however, we found that trials with relatively short duration of follow-up yielded similar results to those with the longest follow-up times. Second, the number of trials with prostate, genitourinary, melanoma and gastric cancers was limited and long-term follow-up of the other large trials should include these as outcomes. Third, the confidence intervals in our metaregression analyses were wide; therefore, small but possibly important differential sub-group effects may have been missed. Fourth, the results of the metaanalyses are dependent on the robustness of the included studies. Some of the randomized trials did not always state whether patients with a past history of cancer were excluded. For example, one trial did not exclude preexisting cancers45, 46 and of the 12 reported cases of breast cancer in pravastatin users in the trial reported by Sacks et al.,41 3 patients had a previous history of breast cancer and one had only been on pravastatin for 6 weeks before being diagnosed. Despite this problem and its inconsistency with the totality of the evidence reported here, the Sacks study is often cited as evidence for a carcinogenic effect of pravastatin.10 Finally, as the primary outcomes of the trials were cardiovascular endpoints, not cancer, and observational studies allow multiple hypothesis testing, there could be unpublished studies with relevant cancer data, leading to publication bias. Investigators are probably more likely, however, not to publish null results and to publish positive or inverse associations. Thus the direction of any bias would be to underestimate the precision (confidence interval) of the null effects we report here, and would not change our conclusions. In support of this argument, Begg and Egger tests suggested no strong evidence of publication bias. Our search included databases logging abstracts of conferences and meetings. These did not yield any trial abstracts that had not yet been fully published, but we did find 11 abstracts of observational studies (9 abstracts of different outcomes from one study72 and 2 reporting on melanoma) without full publications. However, our a priori criteria meant that these were excluded on quality grounds (Fig. 2), as they had not been subject to the levels of scientific peer review applied to full papers.
We conclude that there is no evidence from randomized trials that statin use is associated with cancer risk in the short term. Uncertainty remains, however, for a number of reasons: first, the studies (particularly trials) had relatively short follow-up, so may be too brief to have captured a true association between statin use and cancer incidence or mortality23; second, the observational studies provided weak evidence that statins may have a small protective effect against colorectal cancer and are suggestive of a possible small adverse effect of pravastatin on breast cancer risk; finally, most studies report only on a limited range of cancer sites. Thus, it will be important to continue to follow-up studies, particularly trials, to detect any long-latency effects of statins, to explore possible sources of heterogeneity in metaregression analyses (e.g. by statin type) and to include less common cancer outcomes.
We thank Roger Harbord and David Gunnell for their comments on an earlier draft. Jan Hill obtained references and managed the reference database. We thank Drs. Eric Jacobs and Heather Eliassen for providing additional data on their studies.
Table . Appendix: Medline Search Strategy
Searches were restricted to human populations.
exp lovastatin/or exp simvastatin/ or exp pravastatin/
exp HMG CoA reductase inhibitor/ statins.mp or lovastatin.mp or simvastatin.mp or cerivastatin.mp or fluvastatin.mp or pravastatin.mp or atorvastatin.mp
HMG-CoA reductase inhibitor.mp 3-hydroxy-3-methylglutaryl coenzyme A inhibitor.mp