Dietary Vitamin K Intake and the Risk of Pancreatic Cancer

A Prospective Study of 101,695 American Adults

Dao-Wu Yu; Qu-Jin Li; Long Cheng; Peng-Fei Yang; Wei-Ping Sun; Yang Peng; Jie-Jun Hu; Jing-Jing Wu; Jian-Ping Gong; Guo-Chao Zhong


Am J Epidemiol. 2021;190(10):2029-2041. 

In This Article


We reported the results of this study following the Strengthening the Reporting of Observational Studies in Epidemiology statement.[20]

Study Population

The Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial is a multicenter randomized controlled study designed to determine the potential benefits of selected screening examinations in reducing the risk of mortality from prostate, lung, colorectal, and ovarian cancers. The study design of this trial has been described in detail previously.[21] Briefly, individuals between the ages of 55 and 74 years were recruited to take part in the PLCO Cancer Screening Trial between November 1993 and September 2001. After exclusions (see Figure 1 for the relevant exclusion criteria), approximately 155,000 eligible individuals were enrolled and randomly divided into the intervention arm or the control arm in equal proportions. All individuals provided written informed consent. The PLCO Cancer Screening Trial was approved by the institutional review committees of each screening center and the US National Cancer Institute.

Figure 1.

The study flow chart of identifying eligible subjects, the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial, 1993–2009. The total numbers were not available for the first 2 boxes. DHQ, diet history questionnaire.

The following individuals were further excluded from this study: 1) those who did not complete a diet history questionnaire (DHQ) (n = 36,076); 2) those with an invalid DHQ (missing the date of completion; participant died before the completion date; extreme values of calorie intake (top 1% or bottom 1%); and ≥8 missing frequency responses (n = 5,364)); 3) those who received a diagnosis of cancer before completing a DHQ (n = 9,684); 4) those who did not complete a baseline questionnaire (n = 1940); 5) those diagnosed with pancreatic cancer after the diagnosis of other cancers (n = 42); 6) those diagnosed with endocrine pancreatic cancer (n = 18); and 8) those with outcome events (i.e., incident pancreatic cancer, death, or loss to follow-up) that occurred between trial entry and DHQ completion (n = 68). Finally, a total of 101,695 individuals (49,455 men and 52,240 women) were included (Figure 1). Of note, for most sociodemographic characteristics, the standardized differences between included and excluded populations were found to be less than 0.1,[22] which suggests that the possibility of nonparticipation bias due to the exclusion of a large number of individuals was small in this study (Web Table 1, available at

Assessment of Exposure and Covariates

Dietary intakes of phylloquinone, menaquinones, and dihydrophylloquinone were assessed with a DHQ (version 1.0, National Cancer Institute, 2007; The DHQ is a 137-item self-administered food frequency questionnaire developed to evaluate the frequency and serving sizes of food consumption over the past year. Its performance had been validated in a nationally representative sample of 1,640 individuals; the relevant results suggested that the DHQ had good performance in collecting dietary data.[23] The frequency of consumption and serving size of each food were collected with the DHQ, and then daily food consumption was calculated by multiplying food frequency by serving size for each participant. Dietary intakes of nutrients (including phylloquinone, menaquinone-4 through menaquinone-14, and dihydrophylloquinone) and energy were calculated using the DietCalc software (National Cancer Institute, Bethesda, Maryland),[24] which predominantly considered daily food consumption and nutrient values per 100 grams of food that were from 2 nutrient databases (United States Department of Agriculture's 1994–96 Continuing Survey of Food Intakes by Individuals[25] and the Nutrition Data Systems for Research[26]). Importantly, in this study, dietary intakes involved in all data analyses were adjusted for energy intake from diet using the residual method;[27] dietary menaquinone intake referred to the sum of menaquinone-4 through menaquinone-14; and supplemental vitamin K was not considered, as these data were not available in the PLCO Cancer Screening Trial.

Covariate data, including sex, body weight, height, smoking status, history of diabetes, family history of pancreatic cancer, and aspirin use, were collected with a self-administrated baseline questionnaire. Body mass index (BMI) was computed as body weight in kilograms divided by height in meters squared. Age at DHQ completion, alcohol consumption, and single or multivitamin supplement use data were collected with the above-mentioned DHQ. In this study, physical activity level referred to total time of moderate-to-vigorous activity per and was assessed via a self-administrated supplemental questionnaire. Healthy Eating Index-2015 score was calculated using the method described previously,[28] with higher values suggesting better diet quality.

Assessment of Pancreatic Cancer

Pancreatic cancer was confirmed primarily through an annual study update form that asked individuals to answer whether they received a diagnosis of pancreatic cancer and, if so, the location of the diagnosis and the date of diagnosis. Individuals who did not return the form were contacted repeatedly via e-mail or telephone. Any available medical records were extracted for the further ascertainment of cases of pancreatic cancer reported on the annual study update form. Moreover, death certificates and family reports were used as supplementary means of confirming pancreatic cancer. Notably, to harmonize pancreatic cancer cases, only patients with primary exocrine pancreatic adenocarcinoma (International Classification of Diseases for Oncology, Second Edition codes: C25.0–C25.3 and C25.7–C25.9) were included in our study.

Data Analysis

Because 7 covariates were found to have missing values (Web Table 2), we imputed these missing data using the following procedures to increase statistical power and decrease potential biases. Specifically, for the variable "physical activity level," which had more than 25% missing values, we used multiple imputation with chained equations to impute missing data and set the number of imputations as 25, with the assumption that the data were missing at random;[29] all variables used in data analysis were used to yield the imputed data sets. For the variables with less than 5% missing values, we used the median and the modal value to impute missing data of continuous and categorial variables, respectively. Main data analyses were repeated in individuals with complete data to examine the potential impacts of data imputation on our results.

Cox proportional hazards regression was applied to calculate hazard ratios and 95% confidence intervals to evaluate the association of dietary vitamin K intake with the risk of pancreatic cancer, with follow-up time as time variable. Follow-up time for each participant was calculated as the difference between the DHQ completion date and the pancreatic cancer diagnosis date, death date, loss to follow-up, or end of follow-up (December 31, 2009), whichever occurred first (Figure 2). The proportional hazards assumption was tested through the examination of Schoenfeld residuals;[30] the results suggested that this assumption was satisfied. In regression models, dietary vitamin K intake was classified into quartiles, with the first quartile as the reference group. To test linear trends in hazard ratios across quartiles of dietary vitamin K intake, the median of each quartile was assigned to each participant in the quartile at first to produce an ordinal variable, which was then regarded as a continuous variable in regression models for testing its significance. Covariate selection for multivariable regression models was based on our knowledge of existing literature and the change-in-estimate strategy.[31] Specifically, model 1 (the base model) was adjusted for age and sex; model 2 was further adjusted for factors with a well-established associatoin with the risk of pancreatic cancer, namely smoking status, alcohol consumption, physical activity level, BMI, aspirin use, history of diabetes, and family history of pancreatic cancer; and model 3 was further adjusted for variables selected by the change-in-estimate strategy (i.e., intakes of dietary fiber and folate) and energy intake from diet.

Figure 2.

The timeline and follow-up scheme, the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial, 1993–2009. Note that the time span between 2 events was the average value of all participants.

Restricted cubic spline functions[32] with knots located at the 10th, 50th, and 90th percentiles were applied to investigate the potential dose–response relationship of dietary vitamin K intake with the risk of pancreatic cancer. For comparison with the results of Cox regression, the median of the first quartile was used as the reference level (53.0 μg/day, 3.4 μg/day, and 1.0 μg/day for dietary intakes of phylloquinone, menaquinones, and dihydrophylloquinone, respectively). A P value for nonlinearity was obtained by testing whether the regression coefficient of the second spline was equal to zero.[32]

Several predefined subgroup analyses were performed to determine whether the observed association of dietary vitamin K intake with the risk of pancreatic cancer was modified by age (≥65 vs. <65 years), sex (male vs. female), BMI (≥25 vs. <25), single or multivitamin supplement use (yes vs. no), aspirin use (yes vs. no), smoking status (current or former vs. never), and Healthy Eating Index-2015 score (greater than or equal to the median vs. less than the median). A P value for interaction was obtained by comparing models with and without multiplicative interaction terms before conducting the above subgroup analyses to avoid the potentially spurious subgroup differences.

A series of sensitivity analyses were performed to examine the robustness of our results. The first excluded individuals with extreme values of calorie intake, which refers to <800 or >4,000 kcal/day for men and <500 or >3,500 kcal/day for women.[33] The second excluded individuals with extreme dietary vitamin K intakes (i.e., <5th or >95th percentile). The third excluded individuals who received a diagnosis of pancreatic cancer within the first 2 years of follow-up and the first 5 years of follow-up to assess the potential impact of reverse causation. The fourth used energy-unadjusted dietary vitamin K intake in Cox regression analyses, and the fifth repeated Cox regression analyses with sex-specific quartiles because the distribution of dietary intakes of menaquinones and dihydrophylloquinone differed by sex. The sixth additionally adjusted the fully adjusted model for Healthy Eating Index-2015 score to determine whether the observed associations were mediated by diet quality, and the seventh adjusted for propensity score on the unadjusted model (all covariates in model 3 were used to compute propensity score).

Continuous and categorical variables are shown as mean (standard deviation) and numbers (percentage), respectively. The statistical significance level was set at P < 0.05 under a 2-tailed test. Statistical analyses were performed using STATA, version 12.0 (StataCorp LP, College Station, Texas).