Author + information
- Received March 4, 2008
- Revision received May 27, 2008
- Accepted June 6, 2008
- Published online September 30, 2008.
- Alawi A. Alsheikh-Ali, MD⁎,†,
- Thomas A. Trikalinos, MD⁎,
- David M. Kent, MD, MS⁎ and
- Richard H. Karas, MD, PhD†,⁎ ()
- ↵⁎Reprint requests and correspondence:
Dr. Richard H. Karas, Molecular Cardiology Research Institute, Box # 80, Tufts Medical Center, 800 Washington Street, Boston, Massachusetts 02111
Objectives We sought to assess whether statin-mediated reductions in low-density lipoprotein cholesterol (LDL-C) are associated with an increased risk of cancer.
Background We recently reported an inverse association between on-treatment LDL-C levels and incident cancer in statin-treated patients enrolled in large randomized controlled trials, raising concern that LDL-C lowering by statins may increase cancer risk. However, meta-analyses suggest a neutral overall effect of statins on incident cancer.
Methods A systematic literature search identified 15 eligible randomized controlled trials of statins with ≥1,000 person-years of follow-up that provided on-treatment LDL-C levels and rates of incident cancers (19 statin and 14 control arms, 437,017 person-years cumulative follow-up, and 5,752 incident cancers).
Results In the statin arms, meta-regression analysis demonstrated an inverse association between on-treatment LDL-C and incident cancer, with an excess of 2.2 (95% confidence interval: 0.7 to 3.6) cancers per 1,000 person-years for every 10 mg/dl decrement in on-treatment LDL-C (p = 0.006). The corresponding difference among control arms was 1.2 (95% confidence interval: −0.2 to 2.7, p = 0.09). Compared with the control arms, the statin regression line was significantly shifted leftward, such that similar rates of incident cancer were associated with lower on-treatment LDL-C (p < 0.05). Meta-regression demonstrated that statins lack an effect on cancer risk across all levels of on-treatment LDL-C.
Conclusions There is an inverse association between on-treatment LDL-C and incident cancer. However, statins, despite producing marked reductions in LDL-C, are not associated with an increased risk of cancer.
In a recently published analysis, we examined the relationship between the degree of low-density lipoprotein cholesterol (LDL-C) lowering and adverse events in large randomized controlled trials (RCTs) of hydroxymethylglutaryl coenzyme A reductase inhibitors (statins) (1). During the peer review process, we were asked by the editors and reviewers to include cancer in the analysis, and in doing so observed a significant inverse relationship between on-treatment levels of LDL-C and newly diagnosed cancer in statin-treated patients (1,2). The observed relationship in statin RCTs is consistent with prior epidemiologic observations of an inverse association between serum cholesterol levels and incident cancer (3). These findings raised concern that lowering LDL-C with statins may decrease the risk of cardiovascular events at the expense of an increased risk of cancer. In contrast to these findings, meta-analyses of randomized statin trials have not detected any clinically meaningful effect of statins on cancer incidence (4–6).
In the present study, we examine this apparent paradox by simultaneously considering potential effects of statin use and on-treatment LDL-C levels on cancer risk. First, we expand upon the previously reported association between incident cancers and low on-treatment LDL-C levels in statin-treated patients by accounting for potential confounders in multivariable analyses. Second, we examine the relationship between on-treatment LDL-C and incident cancer in the control arms of statin RCTs and compare this relationship with that found in the statin-treated patients. Finally, we perform a meta-analysis/meta-regression of the effects of statin versus control therapy on cancer incidence, while accounting for on-treatment LDL-C levels.
Eligible studies were RCTs of statin therapy published in the English literature with at least 1,000 person-years of follow-up that reported both on-treatment LDL-C levels and incidence of newly diagnosed cancer during the study follow-up period. We performed a MEDLINE search to identify potential trials published up to July 2007 that would meet our inclusion criteria. Our electronic search strategy included the following terms: hydroxymethylglutaryl-CoA reductase inhibitors, statin, lovastatin, simvastatin, pravastatin, atorvastatin, cerivastatin, fluvastatin, or rosuvastatin. Citations were limited using the terms human, English language, and randomized controlled trial. Additionally, we reviewed the reference lists of published meta-analyses of statin trials to ensure that all appropriate trials were included (4–6).
For each eligible study, the following variables were extracted from the published manuscript: the specific statin used and dose, number of patients in the statin and control arms, duration of follow-up, baseline and on-treatment serum LDL-C levels, and number of patients with newly diagnosed cancer over the period of follow-up. Since nonmelanoma skin cancers were not consistently recorded in all trials, these were not included in the present analysis. In addition, for each trial arm, the following baseline patient characteristics were recorded: age, gender, smoking status, history of diabetes mellitus or hypertension, and body mass index. For categorical variables (e.g., hypertension, smoking status) the proportion of patients with such characteristics at baseline was recorded. For continuous variables (e.g., age, body mass index), the mean or median (as reported in the published manuscript) was recorded. Person-years of follow-up for each study arm were calculated by multiplying the reported follow-up in years by the number of persons in each arm. Risk of incident cancer was estimated by dividing the number of persons with newly diagnosed cancer over person-years of follow-up, and expressed per 1,000 person-years.
On-treatment LDL-C and incident cancer in the statin arms
The relationship between on-treatment serum LDL-C levels and cancer risk in the statin arms was first assessed using univariable meta-regression (see Statistical Methods section). Similarly, we assessed the relationship between cancer risk and each of the following baseline characteristics: mean age, proportion of male subjects, proportion smoking, proportion with diabetes mellitus, proportion with hypertension, and mean body mass index. Each of the baseline factors that showed a significant univariable association with incident cancer was used to adjust the meta-regression of incident cancer risk and on-treatment LDL-C in a multivariable model.
On-treatment LDL-C and incident cancer in the control arms
To understand the role of statin therapy on the observed association between LDL-C levels and cancer risk, we assessed the same relationship (i.e., on-treatment LDL-C and risk of cancer) in the control arms of the statin trials. This offers an opportunity to observe the association of LDL-C and incident cancer in a population otherwise comparable to the statin arms, except for lack of statin therapy. We then compared the association of on-treatment LDL-C levels with cancer risk in the control arms with that in the statin arms (see Statistical Methods section).
Statin's effect on cancer risk adjusting for on-treatment LDL-C levels
While several meta-analyses have shown no significant effect of statin therapy on risk of cancer, it remains possible that the overall neutral effect observed in the previously published meta-analyses obscures an underlying heterogeneity of statin effect on cancer risk based on levels of on-treatment LDL-C. For example, if statins increase the risk of cancer at very low LDL-C levels but have an antineoplastic effect at higher levels (or vice versa), then one would expect an overall neutral effect in a conventional meta-analysis, unless on-treatment LDL-C levels are accounted for. To directly address this possibility, we performed a meta-analysis/meta-regression, examining the effect of statin therapy across the range of on-treatment LDL-C levels in each arm (see Statistical Methods section).
Meta-Regressions of Statin or Control Arms from Large RCTs
In the main analyses, we assumed that incident cancer rates per trial arm were normally distributed. We used random effects meta-regressions to evaluate the association between incident cancer rates and average on-treatment LDL-C levels or other baseline variables, as described in the previous text (7). These analyses take into account both within- and between-arm variability. We performed further analyses assuming that incident cancers follow a Poisson distribution over the follow-up period of each trial (8). Specifically, we used Poisson regressions with robust standard error estimation, and Bayesian random effects Poisson meta-regressions (9). To contrast the association of on-treatment LDL-C levels to cancer risk in the statin versus control arms, we compared the slopes of the 2 regression lines, and used the Chow test to assess whether regression coefficients estimated over statin arms are equal to the coefficients estimated over control arms (i.e., whether the regression lines were identical or not) (10).
Meta-Analysis Adjusting for On-Treatment LDL-C Levels
We calculated the summary effects of statin versus control on incident cancer rates before and after adjustments for average on-treatment LDL-C levels per trial arm. These analyses were again performed in a meta-regression framework, by including indicator variables describing treatment allocation and parent trial, the average on-treatment LDL-C level in each arm, and treatment by LDL-C level interaction. We also performed the corresponding hierarchical random effects Poisson meta-regression analyses in the Bayesian framework.
Analyses were performed in Intercooled Stata 8.2 (Stata Corp., College Station, Texas). All p values are 2-tailed and considered significant at the 0.05 level. No adjustments for multiple comparisons were performed. Bayesian analyses were performed in OpenBUGS and JAGS 0.98 (Martyn Plummer, 2005) programs that use Markov Chain Monte Carlo to obtain posterior distributions of the modeled parameters (11). Noninformative prior distributions were assigned to the cancer incidence rate and its variance, and the coefficient of average on-treatment LDL-C. After visual assessment for convergence, we based results on 20,000 iterations using 2 Markov Chain Monte Carlo chains, a burn in of 10,000 iterations, and a thinning interval of 5. Because the results of Bayesian analyses are posterior distributions, we describe them using their median and a 95% credibility interval (from the 2.5 to 97.5 percentile). See the Online Appendix for further details.
Our search yielded 2,026 citations, which were screened at the abstract level. Of these, 1,319 citations had <1,000 person-years of follow-up, 235 were not of a statin study, and 433 were not RCTs. Accordingly, 39 full-text manuscripts were retrieved for detailed evaluation, of which 24 were eventually excluded for having less than 1,000 person-years of follow-up or for not reporting cancer incidence. Therefore, a total of 15 statin RCTs were eligible for inclusion in the present analysis (Table 1) (12–27). These included 19 statin treatment arms and 14 control arms. There were a total of 51,797 statin-allocated patients and 45,043 control-allocated patients followed over a mean of 4.4 ± 1.4 years (range of follow-up: 0.9 to 6.1 years). The cumulative exposure was 224,886 person-years in the statin arms and 212,131 person-years in the control arms. A total of 5,752 patients with incident cancer were included. The on-treatment LDL-C levels ranged from 89 to 142 mg/dl in the statin arms and 121 to 192 mg/dl in the control arms. The incidence of newly diagnosed cancer ranged from 3.9 to 26.5 per 1,000 person-years in the statin arms and from 6.0 to 23.7 per 1,000 person-years in the control arms.
On-treatment LDL-C and incident cancer in the statin arms
In univariable meta-regressions, there was a significant inverse relationship between on-treatment LDL-C levels and incident cancer in the statin arms, with 2.2 (95% confidence interval [CI]: 0.7 to 3.6) incident cancers per 1,000 person-years for every 10 mg/dl decrement in LDL-C (p = 0.006) (Fig. 1). In similar univariable analyses in the statin arms, there was a significant relationship between incident cancer and age (p = 0.031) and history of hypertension (p = 0.049). In multivariable models, the association between on-treatment LDL-C and incident cancer remained significant after adjusting for any of the available baseline variables (Table 2). Results from Poisson fixed effects as well as Bayesian random effects regressions yielded consistent findings (incidence rate ratio [IRR]: 1.18 [95% CI: 1.09 to 1.27] per 10 mg/dl decrement in LDL-C, p < 0.001, and IRR: 1.20 [95% credibility interval: 1.08 to 1.33] per 10 mg/dl decrement in LDL-C, respectively).
On-treatment LDL-C and incident cancer in the control arms
In univariable analyses, there was also an inverse relationship between on-treatment LDL-C levels and incident cancer in the control arms of the statin trials, albeit not statistically significant in the linear meta-regressions. We estimated an excess of 1.2 (95% CI: −0.2 to 2.7) incident cancers per 1,000 person-years for every 10 mg/dl decrement in LDL-C (p = 0.09) (Fig. 1). However, the corresponding results were highly statistically significant in the Poisson meta-regressions (IRR: 1.11 [95% CI: 1.06 to 1.16], p < 0.001), and were similarly highly significant in Bayesian analyses (IRR: 1.12 [95% credibility interval: >1.00 to 1.26] per 10 mg/dl decrement of LDL-C). While the slopes of the regression lines associating on-treatment LDL-C and risk of cancer were similar in the statin compared with those in the control arms (p = 0.33), overall, the 2 regression lines were significantly different from each other (Chow test p = 0.049 for linear meta-regressions and p < 0.001 for Poisson meta-regressions) (Fig. 1). As such, compared with control subjects, statin-treated patients had lower levels of LDL-C at similar levels of cancer risk.
The effect of statins on cancer risk adjusting for on-treatment LDL-C levels
In unadjusted meta-analyses, statin use was not associated with an increase in cancer rates compared with that seen in the control arms (incident rate difference 0.0 cancers per 1,000 person-years; 95% CI: −0.7 to 0.7, p = 0.99). The same was true when we adjusted for on-treatment LDL-C in each trial and each arm: the summary incident rate difference was 0.2 cancers per 1,000 person-years (95% CI: −5.1 to 5.6, p = 0.92). The hierarchical Bayesian adjusted meta-analysis yielded very similar findings (IRR: 1.02 with 95% credibility interval: 0.95 to 1.10). The neutral effect of statin therapy on incident cancer was consistent across all levels of on-treatment LDL-C levels (p = 0.73 for the effect of on-treatment LDL-C on statin-associated cancer risk) (Fig. 2).
The current analysis of large statin RCTs demonstrates an inverse association between on-treatment LDL-C levels and incident cancer in statin-treated patients that persisted after accounting for patient age, gender, smoking, diabetes mellitus, hypertension, and body mass index. Similarly, a relationship between on-treatment LDL-C levels and incident cancer was also observed in control populations not treated with statins. Importantly, comparison of the association between on-treatment LDL-C and risk of cancer in the statin-treated versus control patients demonstrated that the statin line was significantly shifted horizontally to the left (Fig. 1). In this way, compared with control patients, statin-treated patients achieve lower levels of LDL-C while maintaining similar risks of cancer. In addition, we expanded on the results of previously published meta-analyses looking for relationships between statin therapy and cancer by adjusting for on-treatment LDL-C levels. We observed that the previously reported overall neutral effect of statins on cancer risk holds true for any given level of on-treatment LDL-C (Fig. 2). Taken together, these findings indicate that despite the observed inverse association between on-treatment LDL-C levels and incident cancer, the LDL-C–lowering effect of statins is not associated with an increased risk of cancer, at least over the time frame evaluated here.
The present analysis offers a novel contribution by examining the association of LDL-C and cancer in control cohorts and comparing it with the one we recently reported in statin-treated patients (1). In doing so, it addresses the apparent paradox between the observed inverse association of LDL-C levels and incident cancer rates and the prior meta-analyses of statin trials. The current findings demonstrate that despite the inverse association between LDL-C levels and incident cancer, lowering of LDL-C with statins (i.e., a leftward shift in the horizontal LDL-C axis) is not accompanied by an increased risk of cancer (i.e., no vertical shift along the cancer incidence axis). This supports the notion that lowering LDL-C with statins does not contribute etiologically to cancer. If lowering LDL-C were causally related to cancer, one would have expected an upward shift in cancer risk along the same regression line in statin-treated patients.
The interpretation of these findings and their implications for understanding the observed inverse association between on-treatment LDL-C levels and incident cancer remain complex, as association studies such as those presented here do little to provide mechanistic insight. The main unanswered question is whether this association reflects an etiologic role for lower LDL-C levels in cancer development. The most straightforward interpretation of our findings supports the conclusion that LDL-C levels do not directly affect cancer risk. One potential explanation for this is that the observed associations occur by chance, though the consistency of reports of significant inverse associations between cholesterol levels and cancer risk from epidemiologic studies makes this seem less likely. However, LDL-C levels might be influenced by an unknown confounder that itself is related to cancer risk, while LDL-C levels per se are not, or, alternatively, statins might lower total LDL-C levels without altering potentially oncogenic LDL sub-particles, and in this manner lower total LDL-C levels without affecting the risk of cancer. The current findings do not, however, entirely exclude the possibility of an etiologic role for lower LDL-C levels on cancer risk. For example, statins' effects on cancer may require longer durations of follow-up than observed in the RCTs examined here, or, alternatively, a pro-neoplastic effect of statin-mediated LDL-C lowering is offset by an antineoplastic effect of statins, which could also result in a neutral effect on cancer risk.
There are a number of limitations inherent in the trial-level data analyses presented here. Use of individual patient data would provide a more robust analysis in a variety of ways. Analysis of differences in LDL-C between patients in the same trial arm might reveal associations that are masked by the use of trial level means. With individual patient data one can better account for potential residual confounding that might remain in the current analysis. Individual patient data would also allow for better standardization of interventions (statin type and dosage) and outcomes (definitions of incident cancer), minimize selective reporting of cancer rates, incorporate unpublished data on longer follow-up, model individual patient risks, and perform time-to-event analyses (28–30). In the absence of individual patient data, the possibility that cancers that developed during the course of the study caused lower LDL-C, rather than vice versa, could not be addressed. An additional weakness of relying on trial level data is the lack of standardized adjudication of newly diagnosed cancer. Indeed, newly diagnosed cancer was not reported in several of the large randomized statin trials including recent trials of intensive statin therapy, introducing the possibility of selection bias. Finally, the location of cancer was not consistently reported across the trials, and, hence, whether statins may affect the risk of specific forms of cancer without affecting overall cancer risk could not be determined.
The previously reported association of low levels of on-treatment LDL-C and incident cancer, confirmed here, is not driven by statins, and statin therapy, despite producing marked reductions in LDL-C, is not associated with an increased risk of cancer. Further studies are needed to validate these findings, particularly with longer durations of follow-up.
For statistical information, please see the online version of this article.
Dr. Alsheikh-Ali is currently a recipient of a faculty development award from Pfizer/Tufts Medical Center. Dr. Kent has received research support from Pfizer. Dr. Karas has received honoraria from Merck and Abbott, and research support from AstraZeneca.
- Abbreviations andAcronyms
- confidence interval
- incidence rate ratio
- low-density lipoprotein cholesterol
- randomized controlled trial
- Received March 4, 2008.
- Revision received May 27, 2008.
- Accepted June 6, 2008.
- American College of Cardiology Foundation
- Alsheikh-Ali A.A.,
- Maddukuri P.V.,
- Han H.,
- Karas R.H.
- DeMaria A.N.,
- Ben-Yehuda O.
- Jacobs D.,
- Blackburn H.,
- Higgins M.,
- et al.
- Spiegelhalter D.J.,
- Abrams K.R.,
- Myles J.P.
- Thomas A.,
- O'Hara B.,
- Ligges U.,
- Sturtz S.
- GISSI Prevenzione Investigators (Gruppo Italiano per lo Studio della Sopravvivenza nell'Infarto Miocardico)
- The ALLHAT Officers and Coordinators for the ALLHAT Collaborative Research Group
- Stewart L.A.,
- Tierney J.F.