Statin Exposure and Pancreatic Cancer Incidence: A Japanese Regional Population-Based Cohort Study, the Shizuoka Study

Pancreatic cancer is the seventh most frequent cause of cancer-associated deaths worldwide. Despite recent progress in diagnostics, surgery, and chemotherapy, the 5-year survival rate of patients with pancreatic cancer is <10% (1). The incidence of pancreatic cancer, particularly among Asian populations, was predicted to approximately double from 2018 to 2040 (2). Therefore, the development of preventive treatments for pancreatic cancer is an urgent and daunting challenge.

Statins are among the most frequently prescribed drugs for the treatment of hyperlipidemia to prevent cardiovascular events (3). The pleiotropic effects of statins include inhibiting inflammation and angiogenesis (4). Such effects were predicted decades ago to prevent the growth of cancers (5), including pancreatic cancer (6–8). Unfortunately, progress in this area was hampered by the inconsistent findings of subsequent clinical investigations (9–18). However, two more recent meta-analyses (19, 20) found that statins may be associated with the reduction of the incidence of pancreatic cancer. Previous Asian cohort studies were limited to selected populations such as patients with type II diabetes [relative risk; 95% confidence interval (CI); 0.55 (0.48–0.63)] (11), leading to diminished generalizability of their findings. Moreover, few randomized controlled trials with sufficient subjects and follow-up times, including those of Asian populations, are available.

To address this gap in our knowledge, we employed the Shizuoka Kokuho Database (SKDB), which is a large community-based database that includes general populations and serves to advance preventive health research (21–23). Here, we used the SKDB to investigate the association between statin use and the incidence of pancreatic cancer among the general Japanese population. To control for confounding, we used the inverse probability of treatment weighting (IPTW) method as our major analytic tool.

The SKDB is the first prefecture-wide, longitudinal dataset, and it includes the records of over 2 million residents of Shizuoka Prefecture (24). The SKDB has served as an important data source for several studies (21–24). In summary, the SKDB comprises a community-based longitudinal cohort, including 2,230,848 individuals [women, n = 1,211,161 (54.3%)] who were all Japanese residents of Shizuoka Prefecture located near the center of Japan (population, approximately 3.6 million). We collected comprehensive individual-linked data from the SKDB and assigned unique identifiers to everyone. The contents of the dataset are as follows: basic information from subscriber lists (age, sex, postal code, observation period, reasons for withdrawal including death), claims data of public health insurance agencies [National Health Insurance (NHI), age < 75 years, and those of the Latter-Stage Elderly Medical Care System (LSEMCS), age ≥ 75 years], and the results of public annual health examinations provided to those ≥40 years (Supplementary Table S1; Supplementary Appendix S1).

The SKDB is an appropriate database for real-world studies because it includes precise information about death and loss to follow-up in its basic resident registration system. Each drug and diagnosis national code corresponds to the codes of the Anatomical Therapeutic Chemical (ATC) classification and the International Classification of Diseases, 10th Revision (ICD-10), respectively.

We conducted a retrospective population-based cohort study using the SKDB. The study period was between April 1, 2012, and September 30, 2018. All participants subscribed to the NHI or the LSEMCS within the study period. Cohort entry was defined as the date of registration with those health insurance agencies or April 1, 2012, whichever occurred later. The “index date” was defined as the initial dates of annual health examinations during ≥1.5 years of continuous subscribership after cohort entry. We defined 1.5 years as the “baseline period” before the index date (Supplementary Fig. S1). Demographic information, laboratory measurements, and lifestyle characteristics were collected from the data of the initial annual health examinations on the index dates. The follow-up period was defined as the date of initial diagnosis of pancreatic cancer, the end of the study (September 30, 2018), or withdrawal from the NHI or the LSEMCS, whichever occurred first. The reasons for withdrawal included death, emigration from Shizuoka prefecture, changes of insurer, and other reasons.

We included individuals aged >40 years, and we excluded those who did not undergo annual health examinations after 1.5 years of continuous subscribership, as well as those with histories of any cancer, except for skin cancer other than melanoma (Supplementary Table S2). To specify individuals responsive to statin treatment, we further excluded individuals with high concentrations (≥160 mg/dL) of plasma low-density lipoprotein cholesterol (LDL-c).

We defined statin exposure as ≥1 statin prescription (statins are not commercially available in Japan) during the baseline period, and statin non-exposure as without a statin prescription during the baseline period. We further excluded individuals from the statin exposure group with total statin prescription days <365 during the baseline period to avoid misclassification of statin prescriptions as well as to avoid reverse causation (i.e., symptoms of undiagnosed pancreatic cancer may affect the prescription of statins). We applied an intention-to-treat approach, assuming that statin exposure was unchanged throughout follow-up. The outcome of the study was the incidence of pancreatic cancer, ascertained from claims data using ICD-10 codes (Supplementary Table S2).

We identified potential confounders in previous studies (11, 14, 19, 20, 25, 26) and used them as model covariates. The covariates were as follows: age; sex; body mass index (BMI); current smoking; alcohol consumption; physical activity; health awareness; calendar year of the index date; comorbidities including diabetes, hypertension, smoking-related disease, acute pancreatitis, chronic pancreatitis, cholelithiasis, liver disease, myocardial infarction, peripheral vascular disease, renal disease, and hepatitis B and C virus infection; blood test values including hemoglobin A1c (HbA1c) and estimated glomerular filtration rate; drug history including proton pump inhibitors, nonsteroidal anti-inflammatory drugs (NSAID), hormone replacement therapy, metformin, insulin, dipeptidyl peptidase 4 inhibitors, glucagon-like peptide-1 receptor agonist, warfarin, pancreatic enzymes, and steroids.

To avoid underreporting bias, we further included diabetes and hypertension comorbidities according to drug history and ATC codes. The corresponding ICD-10 codes of comorbidities and ATC codes of drugs are shown in Supplementary Tables S3 and S4, respectively. We classified BMI into four groups according to international criteria (<18.5, 18.5–<25, 25–<30, and ≥30) (27); estimated glomerular filtration rate into four groups according to the criteria for chronic kidney disease (≥60, 45–<60, 30–<45, and <30) (28); and HbA1c into five groups according to a diagnostic criteria of diabetes and the treatment target (<6.5%, 6.5%–<7.0%, 7.0%–<7.5%, 7.5%–<8.0%, and ≥8.0%) (29). We referred to the questionnaires included in annual health examinations to account for current smoking, alcohol consumption, physical activity, and health awareness (Supplementary Table S5).

Categorical values are represented by frequencies and percentages. We estimated propensity scores using logistic regression models with all prespecified covariates to predict the probability of assignment of subjects to the statin-exposure group. To adjust confounders for the relationship between statin exposure and the onset of pancreatic cancer, we employed the IPTW method to calculate the propensity score (30). We assigned patients exposed to statins a weight = 1/(propensity score) and those who were not exposed to statins a weight = 1/(1–propensity score). The strength of the IPTW method is that by increasing the weights for those with unexpected exposures, we can maintain data for all individuals, including the portion of the target population with few observations (31). Thus, we can effectively create a pseudo-population that is balanced across treatment groups. Meanwhile, a problem is the extreme weight for IPTW that arises when propensity score is extremely close to 1 or 0 (32). Therefore, standardized weights (stabilized IPTW) and trimmed weights were employed to deal with the issue of extreme weights (32). We truncated the propensity scores of the statin exposure group at the ≥5th and ≥95th percentiles (0.1–0.66), ensuring that >5% of the number of cases of each group was retained in each layer in a propensity-score interval = 0.05.

We used standardized differences to compare the distributions of baseline characteristics between the two groups before and after adjustment. An absolute standardized difference ≥10% was determined as significant covariate imbalance, and a small value <10% indicates sufficient balance between treatment groups (33). We estimated the hazard ratio (HR) and 95% CI according to the Wald test for the onset of pancreatic cancer with the weighted Cox proportional hazards model.

In the sensitivity analysis, we assessed the robustness of the results using propensity score matching, which was performed using the Greedy nearest neighbor matching without replacement algorithm with one-to-one pair matching. A caliper = 0.2 × standard deviation of the logit of the propensity score was used for calculating propensity score (34). We calculated HR, 95% CI, and corresponding P value according to the Wald test. Furthermore, to determine whether potential unmeasured confounders affected our main outcome, we calculated an E-value (35) that quantifies the required magnitude of an unmeasured confounder that may negate the observed association between statins and incidence of pancreatic cancer.

To evaluate consistency according to prespecified categories (age, sex, history of diabetes and chronic pancreatitis, and LDL-c), we estimated the HRs associated with pancreatic cancer incidence after refitting separate propensity-score weighted survival models for each subgroup.

All analyses followed the principle of intention to treat. All P values are two-sided. Owing to outcomes and subgroups assessed, we report 95% CI and consider P < 0.05 to indicate a significant difference. We used SAS version 9.4 (SAS Institute, Cary, NC) and R version 3.6.2 (The R Foundation for Statistical Computing) for all statistical analyses.

We anonymized the SKDB data (24). The Ethics Committee of Shizuoka General Hospital approved the study protocol (SGHIRB#2020023, 2020).

After applying the inclusion and exclusion criteria, we selected 100,537 individuals for inclusion in the statin-exposure group (24%) and 326,033 individuals for inclusion in the statin nonexposure group (76%; Fig. 1; Table 1). Before adjusting for IPTW, compared with the statin nonexposure group, the statin exposure group was older, included more women, was more health-oriented, had a higher prevalence of comorbidities, and used more drugs other than statins. After adjusting for IPTW, all the covariates were well-balanced between groups.

The person-years of the statin exposure and statin nonexposure groups were 1,098,463 person-years and 352,485 person-years, respectively (Table 2). The details of reasons for withdrawal are shown in Supplementary Table S6. Before adjustment for IPTW, statin exposure was not associated with the reduction of pancreatic cancer incidence (HR, 1.04; 95% CI, 0.93–1.17; P = 0.465). After adjusting for IPTW, the incidence of pancreatic cancer was reduced in the statin exposure group (HR, 0.84; 95% CI, 0.72–0.99; P = 0.036). The E-value of the point estimate = 1.67, and the lower bound of the CI = 1.1. In the sensitivity analysis, a similar result was obtained after adjustment for the propensity matching approach (HR, 0.84; 95% CI, 0.71–0.99; P = 0.041; Table 2, Supplementary Tables S7 and S8). In our prespecified subgroup analyses (Fig. 2), the direction was favorable toward statin exposure within the subgroups of age <70 years, women, patients without diabetes, those without a history of chronic pancreatitis, and LDL-c <120 mg/dL, although the differences were not significant within the subgroups of ages ≥70 years, men, patients with diabetes or history of chronic pancreatitis, and LDL-c ≥120 mg/dL.

Our current results support the conclusion that statin treatment reduces the incidence of pancreatic cancer, and the results are similar to those of the sensitivity analysis. To the best of our knowledge, this is the first study to investigate the relationship between statin exposure and the incidence of pancreatic cancer in the general population of Japanese.

Despite abundant epidemiologic evidence of the effects of statins for preventing pancreatic cancer, the results of real-world studies are inconsistent (9–17). Consistent with our current results, more recent meta-analyses (19, 20) indicate that statin treatment is associated with the reduction of risk of pancreatic cancer in total estimates [pooled odds ratio (OR) = 0.70; 95% CI, 0.60–0.82: summary, relative risk = 0.84; 95% CI, 0.73–0.97, respectively]. However, these findings should be cautiously interpreted considering the profound heterogeneity among studies, mainly attributed to numerous case–control studies [13/27 included studies (19) and 14/26 included studies (20)]. Moreover, Asian populations analyzed in previous epidemiologic studies were selected for individuals with type II diabetes (11) or coronary heart disease (18). The current cohort study may extend previous studies through its investigation of a general Asian population.

General population-based cohort studies conducted in the United States (25, 26) using propensity scores to control for bias did not find a significant association between statin exposure and the incidence of pancreatic cancer. The inconsistent results may be explained by the confounders of pancreatic cancer risk factors (25, 26); e.g., pancreatitis (36), cholelithiasis (37), and the use of antidiabetes drugs (38, 39)]. To the extent possible, here we adjusted for these known environmental risk factors of pancreatic cancer, which we collected from our comprehensive dataset. However, we were unable to collect genetic risk factors such as family histories. Furthermore, there was effect heterogeneity across sex strata in our results, which may be explained that unmeasured confounders were involved in the effect heterogeneity. Regarding the duration of statin use, our study confirmed that there was at least one year of continuous statin use. On the basis of the results of previous studies (11), we believe that one year can be sufficient to demonstrate the benefits of statin use. As for the differences between statin varieties, some previous studies showed that the specific anticancer effects of statins depended on the variety of statin (40), while others showed that there were no differences (10). Thus, further prospective studies or randomized control studies are required, which consider a more comprehensive collection of drug information and known pancreatic cancer risk factors.

In the current study, the use of NSAIDs was prevalent in both groups. One study suggested that the use of NSAIDs alone may reduce the incidence of pancreatic cancer (41), but other studies rejected the association (42). In our study, the amount of NSAIDs was adjusted in a multivariate model. Therefore, the impact of NSAIDs use on the conclusions was considered to be small.

Experimental studies provide evidence supporting the ability of statins to prevent pancreatic cancer. Statin reduces LDL-c receptors by inhibiting 3-hydroxy-3-methylglutaryl coenzyme A reductase, a rate-determining enzyme for the mevalonate pathway. Evidence indicates that the mevalonate pathway produces sterols for membrane structure and nonsterol isoprenoids (43, 44), involving oncogene products such as Ras, Rho, and Rac (4, 45, 46). Moreover, the activity of the mevalonate pathway inhibits the degradation of mutant TP53 (47), a suppressor of human cancers, which is pathologically associated with 75% of pancreatic cancers (48). Furthermore, coenzyme Q, a product of the mevalonate pathway, influences the rate of growth of pancreatic ductal adenocarcinomas in vitro (49). In addition to that, statins are expected to have anticancer effects through off-target effects, including the direct interaction of statins with P-glycoprotein (50).

Whether statins prevent pancreatic cancer in high-risk populations is unknown, despite the findings of studies that investigated the ability of statins to decrease the incidence of pancreatic cancer in high-risk populations. For example, a Danish cohort study of patients with chronic pancreatitis showed no significant association between statin exposure and reduced risk of pancreatic cancer [frequencies: statin users = 0.25% (21/8,278); nonusers = 0.25% (132/52,087); adjusted HR = 0.90; 95% CI, 0.56–1.44] (14). Another Taiwanese cohort study of patients with type II diabetes showed a significant association between statin treatment and a reduced incidence of pancreatic cancer [frequencies: statin users = 0.25% (1,730/690,335), nonusers = 0.14% (611/450,282), adjusted HRs = 0.55; 95% CI, 0.48–0.63] (11). In our subgroup analysis, the preventive effects of statins were less inconsistent in higher risk subgroups (Fig. 2; Supplementary Table S9). This may be explained by the increased opportunity of such high-risk populations to undergo medical examinations such as abdominal echo radiography, leading to the early diagnosis of pancreatic cancer. Therefore, future epidemiologic studies of high-risk populations may require considering this possibility.

The limitations of the current study are mainly associated with its observational nature. First, despite the effort to control for confounding using IPTW, residual and unmeasured confounders can remain because not all confounders were adjusted in the observational study. However, when we calculated the E-value, we found that these confounders may not explain the entire effects of treatment. Second, we were unable to collect information about family history, drug adherence, socioeconomic status, cancer stage, pathologic data, and genetic data. Third, the diagnosis of pancreatic cancer depended on claims data without pathologic information that may lead to differential misclassification. Fourth, we were unable to obtain accurate information about drug doses and the duration of the statin prescriptions, nor were we able to evaluate the dose–response relationship. However, we selected patients who received stains for >365 days. Finally, because this population was exclusively Japanese, our results may not apply to the other races. However, considering the evidence that supports the hypothesis that the antitumor effects of statin may be significant for Asians (19), we believe our results can be generalized to other Asian populations.

In conclusion, our results suggest that statins may suppress the incidence of pancreatic cancer in a general Japanese population. To better understand the associations among high-risk populations, further prospective studies are required to control bias introduced by the increase in medical examinations along with the administration of statins.

Y. Miyachi reports personal fees from Sun Pharma Japan outside the submitted work. H. Itoh reports grants from Shionogi Administration Service Company, grants and personal fees from Takeda Pharmaceutical Company, personal fees from Daiich Sankyo company and Mitsubishi Tanabe Pharma, grants from Mochida Pharmaceutical Corporation and Astellas Pharma, nonfinancial support from Nipro Pharma Corporation, and personal fees from Merck Sharp & Dohme outside the submitted work. No disclosures were reported by the other authors.

K. Saito: Conceptualization, methodology, writing–original draft, writing–review and editing. Y. Sato: Conceptualization, data curation, software, formal analysis, methodology, writing–review and editing. E. Nakatani: Conceptualization, resources, data curation, software, formal analysis, methodology, writing–review and editing. H. Kaneda: Writing–review and editing. S. Yamamoto: Writing–review and editing. Y. Miyachi: Writing–review and editing. H. Itoh: Conceptualization, methodology, writing–review and editing.

We thank Dr. Yasuharu Tabara (Graduate School of Public Health, Shizuoka Graduate University of Public Health), Fumihiro Makita, Kumiko Watanabe, and Naoya Shiotsu (Shizuoka General Hospital) for data management and curating the SKDB data; and Edanz Group (https://en-author-services.edanz.com/ac) for editing a draft of this manuscript.

The Research Support Center in Shizuoka General Hospital conducts contract research projects for public health in Shizuoka Prefecture, funding from Shizuoka Prefecture, including the current study.

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.