Differential and Joint Effects of Metformin and Statins on Overall Survival of Elderly Patients with Pancreatic Adenocarcinoma: A Large Population-Based Study Free

This article is featured in Highlights of This Issue, p. 1161

Pancreatic cancer poses a high mortality burden in the United States, and current therapies have modestly improved survival (1–6). The impetus for exploring chronic disease medications for use in anticancer approaches is to advance new therapeutic strategies into clinical settings with relatively short regulatory evaluation timelines. Among medications being evaluated, metformin and statins are two frequently prescribed drugs with an established safety profile that show promising anticancer effects (7, 8), although their efficacy in pancreatic cancer treatment remains unclear.

Beneficial effects of metformin and statins on pancreatic cancer treatment are biologically plausible (9–13), although findings from epidemiologic studies on their therapeutic benefits have been inconsistent. Our recent meta-analysis included 10 publications on metformin treatment and six publications on statin treatment and found that each medication was associated with improved overall survival in pancreatic cancer patients (14). However, some critical issues have not been addressed. First, dyslipidemia is more prevalent in diabetic patients, and the majority of metformin users also take lipid-lowering drugs such as statins (15–17). Therefore, the survival benefits observed in metformin users may in part be due to the statin use. Moreover, there are few studies that examined the interactive or potential joint effect of combination treatments of metformin and statins (18, 19). Second, misclassification or disregard of immortal time (period of cohort entry and date of first exposure to a drug during which death or an outcome under study could not occur) may introduce unintended bias, which leads to the overestimation of drug effects (20). Several retrospective cohort studies were likely to have suffered from this bias because use of metformin or statin was dichotomized into ever/never categories without taking into account timing of drug use in relation to cohort entry (18, 21–23). Third, previous studies were unable to distinguish the effects of prediagnosis drug use (medications initiated before diagnosis) versus postdiagnosis drug use (medications initiated after diagnosis), and which have been demonstrated in recent studies, such as those using data from the Women's Health Initiative (24). Previous epidemiologic studies on metformin and statins include mostly prediagnosis users, and very few include postdiagnosis users. To address the “healthy user effect” (25), the prediagnosis users should be analyzed separately from postdiagnosis users. Finally, questions regarding how drug initiation and dosages impact on mortality among pancreatic cancer patients remain unanswered. It is therefore important to examine the treatment effect of optimal therapeutic regimens in observational studies.

To inform future trial design and effective clinical practice, these critical issues were examined in a large cohort population with more rigorous approaches. We used a nationally representative cancer database in the United States to determine the differential and joint effects of metformin and statin use on the survival outcomes among elderly pancreatic cancer patients.

After approval by the Institutional Review Board at Rutgers University, this study was performed using the Surveillance, Epidemiology, and End Results (SEER-18) registry linked to Medicare claims. Approximately 97% of U.S. persons ages 65 years or older are eligible for Medicare Part A coverage, which includes hospital, skilled-nursing facility, hospice, and some home health care. Ninety-six percent of elderly Part A beneficiaries chose to enroll in Medicare Part B that covers physician and outpatient services. As of 2016, SEER-Medicare Part A and B data were available for patients diagnosed with cancer through 2013, and data on Medicare Part D for outpatient prescription-drug coverage were available from 2007 to 2012. To allow a 1-year window before or after cancer diagnosis for the baseline assessment of comorbidities and use of prescriptions, we selected patients with primary pancreatic ductal adenocarcinoma (PDAC) diagnosed from January 2008 to December 2011. Primary PDAC cases were identified by using the International Classification of Disease for Oncology, Third Edition (ICD-O-3) histology codes: 8000, 8010, 8020, 8021, 8022, 8140, 8141, 8211, 8230, 8500, 8521, 8050, 8260, 8441, 8450, 8453, 8470, 8471, 8472, 8473, 8480, 8481, 8503 (https://seer.cancer.gov/icd-o-3/). The detailed patient selection steps are illustrated in Fig. 1. Briefly, we excluded patients with nonprimary pancreatic cancer or nonadenocarcinoma histology, or patients in health-care maintenance organizations (HMO). Patients diagnosed at autopsy or with missing diagnosis date or death date equal to or less than diagnosis date were also excluded from the study cohort. To include comorbidity characteristics before cancer diagnosis and treatment characteristics present in inpatient and outpatient claim files, we restricted the analysis dataset to patients who were continuously enrolled in Medicare Part A and Part B from 12 months before cancer diagnosis until death or December 30, 2012, the last date of available Medicare claims data. To capture the potential drug effect up to 3 months before cancer diagnosis and the outcomes after cancer diagnosis, we further restricted our dataset to patients who were continuously enrolled in Medicare Part D beginning 3 months before cancer diagnosis to death or the end of follow-up.

Sociodemographic information, including age, sex, race, and neighborhood income, were obtained from the SEER-Medicare linked databases. We controlled the burden of comorbidities, including cardiovascular disease and chronic kidney disease, using the Deyo adaption of Charlson comorbidity index (26–28). The Charlson comorbidity conditions were identified using inpatient and outpatient hospital claims (Medicare Provider Analysis and Review, Outpatient Standard Analytical File) as well as claim files by individual physicians (Carrier File; refs. 28, 29). To avoid overestimating comorbidity related to PDAC diagnosis, we used the claims in the 11 months prior to cancer diagnosis to calculate the Charlson score. The diabetic comorbidity severity index (DCSI) of diabetic or impaired glucose tolerance (IGT) patients was used to adjust for diabetic severity for predicting mortality (30, 31). Because of the lack of smoking information, we used chronic obstructive pulmonary disease (COPD) as a surrogate for smoking-related illness (32). Tumor characteristics, including tumor grade, stage, and size, were ascertained from the SEER registry database. Treatment information, including resection, radiation, and chemotherapy, was identified from the Medicare Claims (Medicare Provider Analysis and Review, Carrier, Outpatient Standard Analytical File) using the International Classification of Disease, 9th Revision (ICD-9) and the Healthcare Common Procedure Coding System codes (33).

Metformin and statin regimens were extracted from the Medicare Part D Claims file. Daily intensity for metformin was estimated as total milligrams dispensed divided by total days of prescriptions before cancer diagnosis. Statins were categorized based on drug characteristics as: lipophilic (atorvastatin, fluvastatin, lovastatin, and simvastatin); hydrophilic (pravastatin and rosuvastatin); high potency (atorvastatin, rosuvastatin, and simvastatin); or low potency (fluvastatin, lovastatin, and pravastatin) (34). Statin intensity was assessed as an ordinal variable (high, moderate, and low) according to average daily dose lowering of low-density lipoprotein cholesterol (35).

Demographic and clinical characteristics, along with comorbid conditions, were compared among four groups: metformin alone, statin alone, dual users, and neither users using χ² tests. Overall PDAC survival was assessed with survival times calculated as the time from diagnosis to death, or censored at the end of follow-up. Propensity scores (36) were used to estimate the probability of one of the following four exclusive medication use categories: “ever used statin” (p1), “ever used metformin” (p2), “ever used both statin and metformin” (p3), and “none.” They were calculated on the basis of patients' sociodemographic characteristics, and comorbidities (Charlson Comorbidity Score, obesity, chronic pancreatitis, dyslipidemia, and diabetes/IGT) using a general (polytomous) logistic regression model (37). To control pretreatment imbalances on observed variables, propensity scores p1, p2, and p3 were included as covariates in the statistical models for propensity score adjustment.

We first examined the association using a conventional Cox model with medication use as a non–time-dependent variable. To reduce immortal time bias, we then applied a time-dependent covariate Cox regression analysis by treating the medication use as time-varying covariates. The associations between metformin and/or statin use and overall survival were evaluated by sequentially adding the following variables: (i) demographic characteristics, (ii) tumor characteristics, (iii) treatment characteristics, (iv) comorbidities, and (v) Charlson comorbidity score. We developed the final model by adjusting for tumor characteristics, treatment characteristics, and propensity scores, as well as the variables that remained imbalanced after propensity score adjustment (i.e., dyslipidemia, diabetes/IGT, Charlson Comorbidity Score).

To assess the possible effect modification by diabetes/IGT, dyslipidemia, or tumor characteristics, we performed stratified analyses by diabetes/IGT and dyslipidemia as well as tumor stage, grade, and tumor size. In addition, we determined the differential effects on survival for prediagnosis users versus postdiagnosis users, by stratified medication use before and after cancer diagnosis. Furthermore, subset analyses were conducted to assess the effects of metformin use (timing initiation and daily intensity) and statin use (timing initiation, name, type, potency, and intensity) among patients who initiated medications before cancer diagnosis. When we assessed the metformin effect on overall survival, statin use was adjusted as a time-varying covariate. Similarly, we adjusted metformin use to assess statin effect. All statistical analyses were performed using SAS Version 9.4 (SAS Institute).

A total of 12,572 patients with primary PDAC composed the analytic population for the study (Fig. 1). Of 12,572 PDAC patients (median age = 76 years), 11,526 (91.77%) patients died before December 30, 2012, with a median follow-up period of 3.80 months (interquartile range: 1.4–10.1 months). Among them, 950 (7.56%) had used metformin without a statin (metformin alone), 4,506 (35.84%) had used statin without metformin (statin alone), and 2,445 (19.45%) were dual users (Table 1). Compared with Kaplan–Meier median survival of neither users (3.60 months), survival was 3.97, 4.06, and 4.23 months for metformin users, statin users, and dual users, respectively.

Before adjustment using the propensity score, there were significant differences on almost all studied variables among the four groups (Table 1). Metformin alone, statin alone, and dual users were significantly younger than neither users. Statin alone and dual users had a higher proportion of males, lower localized/regional tumor stage, and smaller tumor size, compared with neither users. In addition, metformin alone had higher proportion of chemotherapy and radiotherapy than neither users. Not surprisingly, a higher frequency of patients with dyslipidemia or diabetes/IGT, used either metformin, or statin, or both. After propensity score adjustment, only age, Charlson comorbidity score, dyslipidemia, and diabetes/IGT, could not be well balanced among the four groups (Table 1).

In the conventional Cox model, metformin alone [HR, 0.91; 95% confidence interval (CI), 0.85–0.98)], statin alone (HR, 0.91; 95% CI, 0.88–0.95), or a combination (HR, 0.90; 95% CI, 0.86–0.95) was significantly associated with improved overall survival (Table 2). The estimated HRs for three groups in the time-varying Cox model were larger, compared with those in the conventional Cox model, indicating that immortal time bias had exaggerated medication benefits for cancer patients in the non–time-dependent model. Moreover, only patients on statin alone had significantly reduced overall mortality, and this association remained significant after adjusting for different potential confounders (HR, 0.94; 95% CI, 0.90–0.98). However, the association of overall survival with the exposure to metformin alone or a combination remained nonsignificant (Table 2). When we restricted our population to patients who had survived for greater than 2 months (N = 8,274), we observed a weaker and nonsignificant association between statin use and overall survival (HR, 0.96; 95% CI, 0.91–1.01; Supplementary Table S1), suggesting that the contribution of statins to survival may be mitigated with time.

When stratified by the status of diabetes/IGT or dyslipidemia, we observed similar patterns of results, and no evidence that diabetes/IGT or dyslipidemia modify effects of statins (P_interaction > 0.05; Supplementary Table S2). In addition, we performed stratified analyses on tumor stage, grade, and size to confirm the statin effect. Statin use showed a significant association with improved overall survival among patients with distant tumor stage or III/IV tumor grade, but no effect modification by tumor size was observed (Supplementary Table S3). When we stratified medication use before and after cancer diagnosis, the median survival among prediagnosis and postdiagnosis statin users were 4.07 and 10.93 months, respectively, compared with 3.67 months in nonstatin users (Fig. 2). The adjusted HR for postdiagnosis statin users (0.69; 95% CI, 0.56–0.86) was significantly different from prediagnosis statin users (0.94; 95% CI, 0.91–0.98; P < 0.01; Table 3).

Compared with non-metformin users, patients who used metformin either >6 months or ≤6 months before diagnosis did not have longer survival, however, using metformin 1,000 to 1,500 mg/day was significantly associated with an increased overall mortality (HR, 1.10; 95% CI, 1.03–1.17; Table 4). Compared with nonstatin users, survival benefits were observed in patients who started statins >6 months before diagnosis (HR, 0.95; 95% CI, 0.91–0.98), but not those who started statins ≤6 months before diagnosis (HR, 0.96; 95% CI, 0.87–1.07). No significant difference on survival was observed between patients who started statins >6 months before diagnosis and those who started statins ≤6 months before diagnosis (P = 0.71). In addition, compared to nonstatin users, patients who used rosuvastatin or hydrophilic or high potency statins had longer survival (Table 4). Patients with low intensity statin survived longest (HR, 0.92; 95% CI, 0.87–0.98), followed by moderate intensity (HR, 0.95; 95% CI, 0.90–1.00), whereas high-intensity statin use did not show significant improvement of survival (HR, 0.96; 95% CI, 0.92–1.01; Table 4). The trend of HR from low to high statin intensity was borderline significant (P = 0.07).

To our knowledge, this is the largest U.S.-population-based study to examine the differential and joint effects of metformin and statins on survival among pancreatic cancer patients. We found that exposure to statins, rather than metformin, is associated with an improved overall survival of elderly pancreatic cancer patients. In particular, postdiagnosis exposure to statins is associated with a 31% reduction in mortality, and prediagnosis exposure to statins with a 6% reduction in mortality. Furthermore, the effect of rosuvastatin (the statin with the longest half-life) is most pronounced. These new insights provide crucial data for planning randomized clinical trials using statins as an adjuvant treatment of pancreatic cancer.

Findings from previous epidemiologic studies of metformin use and pancreatic cancer survival are conflicting. Possible explanations include failure to consider diabetic severity and comorbidities (23), statin use (38), and immortal time bias (39). When accounting for these critical issues, we observed that metformin use was not significantly associated with overall survival, either in prediagnosis users or in postdiagnosis users. Our results are consistent with a recent retrospective cohort study of 980 PDAC patients with diabetes when the analysis was performed by using time-varying Cox model (40). These results suggested that the metformin exposure variable is better treated as a time-dependent variable rather than a fixed-time variable (41). In this current study, we acknowledge that we may not have had sufficient power to detect smaller effects due to the limited sample sizes in our subgroup analysis: only 497 patients started metformin ≤6 months before diagnosis, and 700 patients used less than 1,000 mg/day of metformin.

We observed that statin use was significantly associated with improved overall survival of elderly PDAC patients, and this finding was more pronounced in postdiagnosis statin users. A retrospective study that included 1,761 newly diagnosed PDAC patients found that statin use was significantly associated with a lower mortality using a time-dependent Cox model (HR, 0.78; 95% CI, 0.62–0.99; ref. 42). Our results are comparable with the results from a recent SEER-Medicare study of 7,813 elderly PDAC patients (34). The limitations of the previous SEER-Medicare study are the failure to control for important comorbidities and metformin use, as well as the limited sample size. On the basis of previous literature, our recent meta-analysis also suggested that statin use was significantly associated with improved survival of pancreatic cancer patients (HR, 0.75; 95% CI, 0.59–0.90; ref. 14).

Anticancer effects of statins have been demonstrated in various preclinical and mechanistic studies. The main effect of statins is to inhibit cholesterol synthesis through inhibition of the rate-limiting enzyme HMG-CoA reductase (43). In addition, statins have cancer chemotherapeutic properties through various mechanisms: halting cell-cycle progression and proliferation (44); increasing radiosensitization in cancer cells (45); promoting apoptosis (46, 47); and impairing metastasis of tumors (48). On the basis of our current understanding of the diverse molecular pathways of statin action, a protective effect of statins on overall survival of pancreatic cancer patients is biologically plausible. Unexpectedly, this study did not observe a survival benefit for dual users. This might be explained by a higher Charlson comorbidity score and a higher proportion of obesity and diabetes/IGT in dual users, compare to patients with statin alone. The potential antitumor benefits of statins should be carefully further assessed in preclinical and clinical studies to repurpose statins for the treatment of pancreatic cancer.

We observed that that rosuvastatin significantly improved overall survival, whereas other statins did not. Although the mechanisms are unknown, it is worth noting that the half-life elimination of rosuvastatin is much longer (about 20 hours), compared to a half-life of all other statins (which range from 2 and 3 hours). Interestingly, it has been showed that mice xenotransplanted with pancreatic cancer cells and treated with rosuvastatin had higher survival rates, compared with the mice treated with other commercially available statins (49). In addition, we found that hydrophilic and high potency statins exerted better survival outcomes than nonstatin use. These may be due to many factors such as different chemical structures leading to changes in their bioavailability, pharmacokinetics, and pharmacodynamics (50).

We also found low and moderate intensity of statin use improved survival significantly, but high-intensity statin showed a nonsignificant survival benefit. The marked differences were partially due to lowering LDL-C strengths among different intensities, indicating a contribution of downstream intermediates in cholesterol biosynthesis for growth and viability of pancreatic cells. It is possible that the patients using high-intensity statin might have worse hyperlipidemia, and therefore no survival benefit was observed for patients with high intensity statin. Our findings are comparable with those in a recent Kaiser Permanente South California (KPSC) study by Huang and colleagues (51). They reported a 13% and 12% reduction in mortality during study period among any statin users and among prediagnosis statin users, respectively. Furthermore, they found that simvastatin and atorvastatin were the only two medications that were independently associated with improved survival. However, the sample size of the KPSC study was too small to evaluate the effect of rosuvastatin and postdiagnosis statin use.

Several limitations inherent with our data must be addressed. First, the study is limited by its retrospective design and is not a randomized clinical trial. Drug exposures are dependent on other factors related to elderly patients' baseline health. To address this issue, we used propensity scores to estimate the likelihood that patients would use metformin and/or statin. The propensity scores controlled for selection bias that could occur as a result of imbalances of comorbidities and sociodemographic factors (36). Second, we have no drug use information prior to age 65, because only patients ages 65 years or older are eligible for Medicare Part D coverage, leading to possible misclassification of ever use of these drugs if prior users did not file medication claims. However, among 3,579 identified PDAC patients with diabetes/IGT, 1,437 (40.2%) patients (471 of metformin only users plus 966 of dual users) had a history of metformin use and 2,142 (59.9%) never used metformin, which is comparable with the proportion of diabetic PDAC patients using metformin in previous studies (38, 40). Third, we were unable to account for the effect of some established determinants of pancreatic cancer survival, such as cholesterol, triglyceride, glucose level, liver-specific metastases, CA 19-9, or CEA. Fourth, we lacked individual-level data on socioeconomic status of cancer patients, which might confound the association of metformin/statin and mortality. Finally, due to data limitations, tumor characteristics such as tumor grade and tumor size were missing on over 20% of the subjects.

Considering these limitations is critical for interpreting the strength of evidence of survival benefit, and future observational studies should focus on addressing the possible effect of timing as well as type and dosage of statin use on pancreatic cancer survival. Further prospective studies with solid rationale for evaluating the use of statin in conjunction with chemotherapy for pancreatic cancer can help determine if statin is an effective treatment for pancreatic cancer.

No potential conflicts of interest were disclosed.

Conception and design: J.-Y. E, S.-E. Lu, Y. Lin, G. Lu-Yao, X.-L. Tan

Development of methodology: J.-Y. E, S.-E. Lu, Y. Lin, K. Demissie, G. Lu-Yao, X.-L. Tan

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): G. Lu-Yao, X.-L. Tan

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.-Y. E, S.-E. Lu, Y. Lin, J.M. Graber, D. Rotter, L. Zhang, K. Demissie, G. Lu-Yao, X.-L. Tan

Writing, review, and/or revision of the manuscript: J.-Y. E, S.-E. Lu, Y. Lin, J.M. Graber, L. Zhang, G.M. Petersen, K. Demissie, G. Lu-Yao, X.-L. Tan

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.-Y. E, J.M. Graber, X.-L. Tan

Study supervision: X.-L. Tan

Other (Consulted in results interpretation and clarification of design): G.M. Petersen

This study used the linked SEER-Medicare database. The interpretation and reporting of these data are the sole responsibility of the authors. The authors acknowledge the efforts of the Surveillance, Epidemiology, and End Results (SEER) Program tumor registries in the creation of the SEER-Medicare database.

X.L. Tan is supported by a grant from the National Cancer Institute at the NIH (K07CA190541). This research was in part supported by the Biometrics Shared Resource of the Rutgers Cancer Institute of New Jersey (P30CA072720).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.