Cancer places one of the biggest burdens on health care system in both highly developed and less developed countries, and its incidence and mortality have been increasing for decades mainly due to longer life expectancy [1
]. Based on cancer statistics, 14.1 million new cancers occurred in 2012, while 8.2 million people died of cancer [2
]. Despite the efforts of industry and physicians, overall survival and progression-free survival is still unsatisfactory for most cancer types.
Statins, competitively inhibiting 3-hydroxy-3-methylglutaryl coenzyme A reductase, have been used for lowering cholesterol levels [3
]. Because of this effect of statin, statins have proven to be effective in reducing the risk of vascular diseases, such as coronary artery disease and stroke [4
]. Besides their lipid-lowering effects, statins also exhibit anti-inflammatory, immunomodulatory and antithrombotic effects [6
]. It has been proposed that statins also have anti-tumor effects. The mechanism of anti-tumor effect is poorly understood, but some in vitro studies indicate that statins suppress proliferation of tumor cells and angiogenesis [9
]. Furthermore, epidemiologic studies, clinical trials and meta-analyses also support this benefit in different types of cancer, yet there is a lack of studies, and conflicting results on the relationship between statin use and cancer incidence [11
To understand and evaluate the strength of the evidence of the effect of statins on reducing cancer incidence, we carried out an umbrella review and comprehensively re-analyzed the data of meta-analyses of randomized controlled trials (RCTs) and observational studies.
2. Materials and Methods
We performed an umbrella review of meta-analyses and systematic reviews reporting on the associations between statin use and the incidence of cancer. This umbrella review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [14
]. The PRISMA checklist is shown in the Supplementary Materials
2.1. Literature Search
We searched the PubMed database and limited the articles to those written in English, regardless of the publication date. The final search was performed in August 2018. The keywords we used were the following: ‘(hydroxymethyl glutaryl-coa reductase inhibitor OR statin) AND (cancer OR neoplasm OR tumor) AND (meta-analysis OR systematic review)’. Meta analyses of either RCTs or observational studies were included in our search strategy. We reviewed the retrieved articles by examining the titles, the abstracts, and the full texts, then decided which article to include or exclude. We further searched the EMBASE database for potentially eligible meta-analyses, but no additional meta-analysis was included because the identified meta-analyses were lacking data necessary for performing re-analysis or overlapped with the PubMed search. The detailed search strategy is presented in Figure 1
2.2. Eligibility Criteria and Data Extraction
We included meta-analyses and systematic reviews of both RCTs and observational studies reporting on the relationship between statin use and cancer incidence. Observational studies included both cohort and case-control studies. We excluded review articles without meta-analysis, in vitro studies, and genetic studies. We also excluded meta-analyses lacking data necessary for performing re-analysis. If an article presented more than one meta-analysis, all meta-analyses were included and assessed separately by study design or cancer type.
Two investigators (G.H.J. and J.I.S) independently extracted the data, and discrepancies were resolved through consensus. We obtained the data from eligible meta-analyses and extracted and summarized the information on first author, year of publication, the type of cancer, the study design, the number of included studies, the number of cancer cases and total participants, and the random effects with a 95% confidence interval (CI). From the eligible studies, we also extracted the raw data of each individual study for further meta-analysis by combining all data by cancer and study design. If a single study consisted of both RCTs and observational studies, we separated the studies according to the study types (RCTs, observational studies, case-control, and cohort) and reported the results separately.
2.3. Statistical Analysis
We firstly re-analyzed each meta-analysis and reported the relationship between statin use and cancer incidence. In addition, if there were overlapping meta-analyses on the same topic, we combined all the individual studies from eligible meta-analyses according to the type of cancer and study design and performed a re-meta-analysis after eliminating overlapping individual studies and including missing individual studies. We presented the summary effect size, 95% CI, and p-value with both random- and fixed-effects. All re-analyses in this study were performed using the Comprehensive Meta-Analysis software ver.3.3.070 (Borestein, NH, USA).
2.4. Estimation of Summary Effects and Estimation of Prediction Interval
For each meta-analysis, we re-analyzed the individual studies and estimated the summary effects and 95% CI using both random- and fixed-effects methods [15
]. We also calculated and presented the 95% prediction interval (PI), which address the dispersion of effects (in 95% of cases the true effect in a new study will fall within the PI) and further account for between-study heterogeneity [16
], whereas CI reflects the accuracy of the mean.
2.5. Evaluation of Between-Study Heterogeneity and Small Study Effects
We assessed heterogeneity across the studies using the I2
metric of inconsistency and the p
value of the X2
-based Cochrane Q test. I2
values of <50%, 50%–75%, and >75% are usually judged to represent low or moderate, large and very large heterogeneity, respectively [17
Publication bias was evaluated by using Egger’s regression test [18
]. Small study effects were used for detecting publication and reporting bias [19
]. When the Egger’s test was significant (p
-value < 0.10) in random-effects meta-analyses, we decided that the study has small-study effects.
2.6. Determination of the Level of Evidence
We determined the level of evidence of each meta-analysis and re-analyzed the pooled meta-analysis to classify the strength of the evidence of the association between statin use and cancer incidence. The criteria were set according to the statistical significance by random and fixed-effects p
-values, 95% PI, a small-study effects, a between-study heterogeneity, and concordance between the effect estimate of the largest study and summary estimate of the meta-analysis [21
]. The criteria were as follows:
Convincing evidence: There was a statistical significance for the random-effect and fixed-effect p-values at p < 0.001. No small study effects or large between-study heterogeneity were found, and 95% PI rejected the null hypothesis. There was a concordance between the effect estimate of the largest study and the summary effect of the random-effects meta-analysis.
Suggestive evidence: There was statistical significance of random effects at p < 0.05, but a 95% PI included the null hypothesis. No small study effects or large between-study heterogeneity were found.
Weak evidence: There was a statistical significance of random effects at p < 0.05. Small study effects or large between-study heterogeneity were found.
Non-significant association: There was no statistical significance by random effect meta-analysis (p > 0.05)
However, if large heterogeneity was found, we rechecked the results to determine whether it might be due to differences in the direction of the effect or if it could be due to differences in the size of the association. In the latter case, we re-determined the level of evidence again.
The purpose of this umbrella review of previous meta-analyses and re-analysis of meta-analyses, including all the individual studies, was to highlight the potential effects of statin use on cancer incidence. We re-analyzed the data from 43 meta-analyses to evaluate the associations between use of statins and cancer incidence. By only using a random-effects p-value, 10 of 18 associations of cancer incidence showed a statistically significant preventive effect of statin.
Although there was a weak or non-significant preventive effect of statin use on most cancer types, there was a suggestive level of evidence regarding the preventive effects of statin use on four cancer types (esophageal cancer, hematological cancer, leukemia, and liver cancer). Re-analysis of association between statin use and leukemia incidence was performed with one eligible meta-analysis [31
] consisting of nine individual studies, which might be relatively a small number of individual studies for re-analysis. However, associations of the other three cancer types had an adequate number of individual studies (27 for esophageal cancer, 34 for hematological cancer, and 27 for liver cancer). A large number of the included studies for meta-analyses are considered to be valid [65
], and, therefore, the outcomes for the 3 cancer types mentioned above might be plausible. Six types of cancer had weak evidence due to substantial publication bias and significant heterogeneity established by the
Although most of the re-analyses showed weak or non-significant evidence, the conventional interpretation of current meta-analysis is that there was preventive effect of statin use on cancer incidence in some cancer types, based on a random effects p
-value, an effect size with 95% CI [66
]. According to these criteria, 10 of 18 meta-analyses on cancer incidence outcomes demonstrated that statins have a preventive effect on cancer risk.
In addition, while most of the statistically significant individual meta-analyses showed that statins have a preventive effect on cancer, one meta-analysis of observational studies on association with statin and non-melanoma skin cancer suggested that there was a positive relationship between statin and non-melanoma skin cancer [24
]. Yang et al. suggested that meta-analyses of observational studies might show more noteworthy result due to the characteristic of observational studies, since it may have advantage of examining rare occurrences of diseases such as cancer. However, the level of evidence in this study was weak, and it included only one meta-analysis. Therefore, we must scrutinize the validity of the results. Further meta-analyses with additional studies will be needed.
Our study evaluates the strength of evidence using multiple values presented or calculated in each meta-analysis. The strength of evidence reinforces the results from the meta-analyses and helps choose the best evidence. Various methods for assessing the evidence level are presented, yet there is no definite grading method for an umbrella review [67
]. Recent umbrella reviews include the p
-value of the meta-analysis, between-study heterogeneity, small study effect, and 95% PI for the grading the level of evidence, which is more related to quantitative values [20
In addition, substantial heterogeneity is an issue in systematic review and meta-analysis. It is essential to explain and manage the heterogeneity to underline the validity of the respective findings [70
]. Umbrella reviews that re-analyze meta-analyses include large number of individual studies, and, therefore, controlling their heterogeneity can be troublesome. Previous umbrella reviews determined a large heterogeneity of
> 50–75%% as weak evidence [71
]. However, this application should be applied cautiously, because heterogeneity can increase if the number of individual studies increases. In addition, if the heterogeneity is large, it can be due to differences in the direction of the effect, or it can be due to differences in the size of the association. In the latter case, therefore, we thought the level of evidence should be re-determined and upgraded the level of evidence from weak to suggestive.
In the eligible meta-analyses, overlapping meta-analyses on the same topic were frequently reported (Table 3
). Overlapping meta-analyses may give an ambiguous result and should be acknowledged [73
]. There are several ways to overcome this problem, and we carried out re-analysis by merging all the extracted individual studies with coherent data. Integration of data from meta-analyses might have more strengths than assembling existing reviews [74
]. In our study, the incidence of lymphoma associated with statin use showed a statistically significant outcome with a weak level of evidence, but two other eligible individual meta-analyses of the same association were not significant (Supplementary Table S1
). Also, re-analysis of the association of incidence of prostate cancer with statin use was graded as weak evidence, but recent meta-analysis performed by Raval et al. [29
] was not significant. Raval et al. only included 27 individual studies, but our study included a total of 64 individual studies, which highlights that there may have been missing eligible studies even in the recent meta-analysis. The comparison of the results of our study and the largest meta-analysis are presented in Table 4
Results of the meta-analysis can be influenced by study design. Aromataris et al. reported that the types of studies should be matched in systematic reviews, and meta-analyses to be considered for its primary objective [75
]. In our study, no meta-analysis that included only RCTs showed a significant preventive effect of statin use on cancer incidence, but re-analyses of observational studies showed statistically significant findings in 11 of the 18 statin–cancer associations. Among these 11 associations, the results of the overall studies (RCTs and observational studies) were determined by those of observational studies in 10 cancers, except non-melanoma skin cancer, for which the results were determined by RCTs. The heterogeneity between overall study design and observational studies may be due to the relatively large number of observational studies included. In addition, observational studies tend to have more biases than RCTs [76
], and some reports suggest that the outcomes of observed associations could be false positives or inflated if a large between-study heterogeneity is present [77
]. However, meta-analysis of RCTs should be interpreted carefully because cancer events are not the primary endpoints of clinical trials. Besides, duration of treatment in clinical trials was relatively shorter than that in observational studies, so there may be an uncertainty of the association.
Our study has several limitations. First, we only assessed individual studies from systematic reviews and meta-analyses eligible for re-analysis, and, therefore, some very recent individual studies might have been missed. However, considering that even a very recent updated meta-analysis for one cancer missed many individual studies, even though thorough search strategy was performed using many search sites such as PubMed, Embase, Scopus, Cochrane database, etc., we think that one should also check the individual studies from previous meta-analyses when updating the meta-analysis. Second, individual studies can have biases, but assessing the quality of individual studies was beyond the scope of our review. Third, exploring the association between dose and types of statins and cancer incidence was also beyond the scope. Likewise, due to a lack of applicable data, we could not stratify the effect of statins by participant age or duration of treatment, which may be the parameter needed to evaluate the true association. Fourth, there were statistical limitations. A 95% PI and Egger p-value could not be assessed if there were only two or fewer individual studies. There were also some missing data in the largest study effect when there was no number of population data in individual studies. Finally, 95% PI, between-study heterogeneity, and publication bias may not be definitive criteria for assessing the strength of the evidence.
Nevertheless, in summary, we extensively re-analyzed meta-analyses on the associations between statin use and cancer incidence. In 10 of 18 studies there were significant relationships between statin use and cancer incidence. Although many meta-analyses of RCTs and observational studies reported significant associations between statin use and cancer incidence, only a small portion of these associations were without biases. Also, there was an individual meta-analysis reporting increased risk of cancer associated with statins use, which should be carefully interpreted by researchers and clinicians. Future studies should include more precise individual data, assessment of potential bias, and updated meta-analyses with more qualified RCTs and observational studies. We suggest that clinicians carefully consider the effects of statins on incidence of different types of cancer on the basis of the findings of our study.