Association between serum retinol and overall and cause-specific mortality in a 30-year prospective cohort study

How retinol as a clinical indicator of vitamin A status is related to long-term mortality is unknown. Here we report the results of a prospective analysis examining associations between serum retinol and risk of overall and cause-specific mortality. During a 30-year cohort follow-up, 23,797 deaths were identified among 29,104 men. Participants with higher serum retinol experienced significantly lower overall, CVD, heart disease, and respiratory disease mortality compared to men with the lowest retinol concentrations, reflecting 17–32% lower mortality risk (Ptrend < 0.0001). The retinol-overall mortality association is similar across subgroups of smoking intensity, alcohol consumption, body mass index, trial supplementation, serum alpha-tocopherol and beta-carotene concentrations, and follow-up time. Mediation analysis indicated that <3% of the effects of smoking duration and diabetes mellitus on mortality were mediated through retinol concentration. These findings indicate higher serum retinol is associated with lower overall mortality, including death from cardiovascular, heart, and respiratory diseases.

V itamin A is a class of essential, fat-soluble compounds of which retinol is the major form in circulation and considered the biochemical indicator of human vitamin A status 1 . Animal products are the primary food sources of retinol, including liver, butter, milk, cheese, and egg. Retinol and related retinoids act as ligands for cellular retinoic acid and retinoid X receptors (RARs and RXRs) through which they modulate transcription of genes regulating a wide-range of physiological processes including embryonic development, post-natal growth, vision, epithelial differentiation and maintenance, and immunity [1][2][3] . The deficiency of vitamin A can lead to xerophthalmia and blindness, infections and inflammation, chronic diseases, and even death. Despite the well-established clinical importance of vitamin A sufficiency for growth and development, physiological homeostasis, and overall human health, a comprehensive assessment of the long-term relationship between vitamin A status and overall and cause-specific mortality is lacking.
In order to more robustly evaluate the vitamin A-mortality relationship, we conducted a prospective analysis of a large cohort of men with serum retinol measured at study entry and over three decades of subsequent observation with nearly 24,000 deaths from heart disease, stroke, other CVD, cancer, respiratory disease, diabetes mellitus, and other causes. Dose-response and population subgroups of lifestyle and other selected factors were also examined. Here, we show statistically significant inverse associations between higher serum retinol and lower risk of overall and cause-specific mortality, including death from cardiovascular, heart, and respiratory diseases. The inverse retinol-overall mortality association is similar across several cohort subgroups, including smoking intensity, alcohol consumption, body mass index, trial supplementation, serum alpha-tocopherol, and betacarotene concentrations, and duration of follow-up.

Results
Baseline characteristics. The median baseline serum retinol concentration in the cohort was 576 µg/L (interquartile range = 500 µg/L to 662 µg/L), and the mean retinol value in the fifth quintile was 84% higher than that in the first quintile. According to the WHO criteria, the present study had only two participants identified with severe vitamin A deficiency (≤98 µg/L), and 21 participants identified with subclinical vitamin A deficiency (98 to ≤196 µg/L). When compared with men in the lowest quintile category, men in higher quintiles of retinol were slightly younger, had higher blood pressure, BMI, and educational level, and were more likely to have a history of CVD, use vitamin supplements and be more physically active (Table 1). Serum retinol was also positively associated with serum total and HDL cholesterol and alcohol consumption (as expected) [12][13][14] , and inversely associated with serum beta-carotene. With regard to the magnitude of the correlation, serum retinol was negligibly or weakly correlated with these factors (all Pearson correlation coefficients, −0.12 < r ≤ 0.22) 15 .
Serum retinol and overall and cause-specific mortality. During 31 years of follow-up (median 18 years), there were 514,289 person-years of accumulated observation, with 23,797 deaths,  including 9,869 deaths from CVD (8,064 heart disease and  1,764 stroke), 7,695 deaths from cancer, 2,161 deaths from  respiratory disease, 119 deaths from diabetes, 1,255 deaths from  injuries and accidents, , for all-causes combined, CVD, heart disease, respiratory disease and other causes, respectively ( Table 3).
The restricted 4-knot cubic spline linear regression of the serum retinol-mortality association substantiated a non-linear dose-response, with all-cause and cause-specific mortality (mortality from CVD and heart disease) increasing among men with serum retinol concentrations below 483 µg/L (cutoff reference value of the lowest quintile) (Fig. 1). Increasing serum retinol was related to decreasing overall, CVD and heart disease mortality, with greatest risk reduction apparent for concentrations near 660 µg/L. The multivariable fractional polynomials analysis, after adjustment for confounding factors, showed similar patterns of the dose-response associations between serum retinol and overall mortality as well as CVD and heart disease mortality (P values < 0.0001 for multivariable fractional polynomials compared with linear, Supplemental Fig. 1). By contrast, serum retinol appeared to be linearly associated with respiratory disease mortality (on the log scale of the hazard ratio) (restricted cubic splines model, P nonlinear = 0.04, Fig. 1; multivariable fractional polynomials model, P nonlinear > 0.05, Supplemental Fig. 1). Kaplan−Meier survival plots show that compared with those in the higher quintiles, men in the lowest category of serum retinol experienced excess cumulative overall mortality and mortality from CVD, heart disease and respiratory disease (all log-rank P values < 0.0001; Fig. 2).
Stratified analyses. Stratified analyses revealed no significant differences in the retinol-overall mortality association across subgroups of smoking intensity, alcohol consumption, BMI, trial supplementation, serum alpha-tocopherol, and beta-carotene concentrations, and duration of follow-up (Fig. 3). There was, however, a significant interaction with both age and serum total cholesterol where the inverse association appeared stronger among men who were older than 54 years and those with low total cholesterol concentrations (P interaction < 0.005, respectively). Findings were similar across subgroups for CVD, heart disease, and respiratory disease mortality with no significant differences observed (Supplemental Tables 1-3).
Sensitivity analyses. The inverse serum retinol associations with mortality from CVD, heart disease, and respiratory disease were accentuated after excluding the 12,488 men with a baseline history of CVD or diabetes mellitus (mortality from CVD, heart disease, and respiratory disease in the multivariable-adjusted model, fifth versus first quintile: HRs = 0.72, 0.68 and 0.60, respectively; all P trend < 0.0001). By contrast, the retinol-overall mortality association remaining unchanged (HR = 0.81, P trend < 0.0001; Supplemental Table 4). The findings were similar after excluding the first 5 years of follow-up (Supplemental Table 5) and when used serum retinol concentration data from the third follow-up year (fifth versus first quintile, HR [95% CIs] for overall mortality: 0.85 (0.81, 0.89), P trend < 0.0001; Supplemental Table 6).
Mediation analyses. As reported, traditional risk factors including obesity, intensity of smoking, duration of smoking, diabetes mellitus, and alcohol drinking were statistically significantly associated with increased risk of overall mortality, whereas light or moderate physical activity was inversely associated risk of mortality in the cohort (Table 4). Our results from the mediation analysis suggest that 2.2% (1.5−3.2) of the smoking duration association with mortality, as well as 1.2% (0.5−3.1) of the diabetes mellitus association, were mediated through serum retinol concentration.

Discussion
In this large prospective serological analysis involving 29,104 cohort participants and 23,797 deaths over three decades, we found significant inverse associations between serum vitamin A (retinol, subsequently referred as "vitamin A") and mortality from overall, CVD, heart disease and respiratory disease after adjusting for several important risk factors, with subjects in the highest serum quintile experiencing 17-32% lower mortality when compared with those in the lowest quintile category (5-12% per one SD increment). The beneficial overall mortality association was evident across all population subgroups with the exception of being stronger among older men and those with lower serum total cholesterol. The effect sizes of these observed risk estimates were relatively small, however.
The present investigation is the largest to examine the association of vitamin A biochemical concentrations with all-cause and cause-specific mortality, and our data suggests mortality declines with increasing retinol concentration, with a risk nadir for overall and CVD mortality in older men with serum values of 600-700 µg/L, and excess mortality among men with retinol below 500 µg/L (n = 7,321, which included two and 21 participants that exhibited severe [≤98 µg/L] or subclinical [98 to ≤196 µg/L] vitamin A deficiency, respectively). Humans cannot synthesize vitamin A and derive this essential nutrient from animal product-based retinyl esters and plant-based pro-vitamin A carotenoids (e.g., alpha-and beta-carotene) 16,17 . For example, the primary dietary sources for dietary vitamin A in this population (as mean percentage of total daily dietary vitamin A) were liver (33.9%), butter (16.4%), egg (9.2%), milk (8.6%), cheese (3.5%) and other food items combined (28.4%). Preformed retinol is absorbed from the intestine, esterified, and transported in chylomicrons to the liver. For healthy, well-nourished individuals, it is estimated that 60-95% of vitamin A is stored in the liver which plays a central role in its metabolism and homeostasis through hydrolysis of stored retinyl esters, complexing of retinol with retinol-binding protein (RBP), and release of this "holo-RBP" into systemic circulation for uptake and use by other    [16][17][18][19] . Previous studies raised the possibility that both insufficient and excess vitamin A status may be associated with increased risk of adverse health outcomes 8,20 , including as a result of direct toxic properties of hypervitaminosis A 17 . We did not observe excess mortality among men with high serum retinol, however, possibly because of a limited range of concentrations in our study population. The threshold of toxic concentrations of retinol, as well as whether there is a positive association of Table 3 Absolute risk differences for the association between serum retinol and overall and cause-specific mortality. mortality risk with excess concentrations of serum retinol, can be evaluated in further studies of populations with available data of excessive serum retinol values. Given the fact that vitamin A is metabolized and stored in the liver 17 , a positive association between liver consumption and serum retinol is biologically plausible. However, along with previous findings [12][13][14]21,22 , our data showed a negligible association between serum retinol and dietary liver consumption (Pearson correlation coefficient = 0.05, P < 0.0001), and a weak association with alcohol consumption (Pearson correlation coefficient = 0.20, P < 0.0001), with negligible and weak correlations being defined as correlation coefficients that range from 0.00 to 0.10 and 0.10 to 0.39, respectively 15 . Additionally, our data showed that the inverse association between serum retinol and risk of overall mortality was not changed by additional adjustment for liver consumption (Q5 versus Q1: HR = 0.84, 95% CI: 0.80, 0.87, P for trend < 0.0001). This suggests that a vitamin A-enriched diet was not sufficient to modify the associations appreciably. Findings from prior studies regarding vitamin A status and overall mortality have been conflicting, with several investigations showing no associations, dose-response or otherwise 4-8 . The NHANES III Study (The Third National Health and Nutrition Examination Survey) of 16,008 participants with 4,225 deaths showed a U-shaped association between serum retinol and total mortality such that individuals with concentrations in either the quintile 3 or 4 category (but not in quintile 5) experienced 18% reduced total mortality compared with those in the lowest quintile (HRs 95% CIs = 0.82 [0.68, 0.98] and 0.82 [0.69, 0.97], respectively; P trend = 0.93) 7 . Using the same NHANES III Study with 6,069 participants aged 50 years or older, a U-shaped overall mortality association was found with HRs of 2.9 and 1.2 for participants with vitamin A status that was deficient (<30 µg/dL) or excessive (>80 µg/dL) when compared to those with normal retinol concentrations (30-80 µg/dL) 8 , respectively. In a cohort of 379 renal transplant recipients with a median of 1,739 days follow-up, serum retinol concentration was related to a decreased risk of all-cause mortality 9 . For established mortality risk factors, our mediation analysis suggests that only a small, <3% role is played by serum retinol for the association of overall mortality with the duration of smoking and diabetes mellitus. Whether retinol mediates portions of other risk factor-mortality associations (e.g., lipid impairment) can be investigated in future studies.
Retinol biochemical status and cause-specific mortality, particularly CVD mortality, has been only sporadically studied, with  [4][5][6][7] . By contrast, serum retinoic acid was inversely associated with overall and CVD mortality among 1,499 patients with coronary artery disease, with those in the highest serum quartile experiencing 35 and 40% lower mortality, respectively, as compared with those in the lowest quartile category 10 . In 1,530 acute ischemic stroke patients, lower retinoic acid concentrations were related to higher risk of overall and CVD mortality, independent of other risk factors 23 , results that were essentially consistent with the present findings. Case-control studies have also documented inverse associations of plasma retinol with risk of CVD including coronary heart disease 13 , stroke 12 , and CVD mortality 24 . Such protective vitamin A-CVD mortality associations are supported by several plausible biological mechanisms. Retinol, its derivate retinoids, and their receptors exhibit potent inhibition of the atherosclerotic disease through cholesterol efflux from macrophages, enhanced reendothelialization, and reduced proinflammatory response. Cholesterol efflux capacity has been characterized as reverse cholesterol transport and shown prospectively to be inversely related to CVD events including mortality [25][26][27] . 9-Cisretinoic acid also suppresses lipoprotein-derived cholesterol accumulation in macrophages and inhibits foam cell formation 28 Fig. 3 Hazard ratios for overall mortality by serum retinol, stratified by selected factors. Hazard ratios of overall mortality were for one SD increment of serum retinol concentration. We used Cox proportional hazard regression models to estimate HRs and 95% CIs. Models were adjusted for age, BMI, serum total and serum HDL cholesterol, cigarettes smoked per day, years of smoking, alcohol intake, intervention assignment, systolic and diastolic blood pressure, history of CVD, and history of diabetes. P value for interaction: according to the likelihood test to assess the statistical significance of the crossproduct term entered into the Cox proportional hazard regression model. Abbreviations: BMI body mass index; CI confidence interval; CVD cardiovascular disease; HDL high-density lipoprotein; HR hazard ratios. *P value for interaction achieved the Bonferroni corrected threshold: 0.05/10 = 0.005. Table 4 Mediating influence of serum retinol on the associations between major risk factors and overall mortality.
In contrast to CVD, there are relatively few studies of circulating vitamin A and respiratory disease mortality. The British National Diet and Nutrition Survey study of 1054 participants observed a non-significant inverse association for respiratory disease mortality with an age-and sex-adjusted 12% lower risk per SD increment in plasma vitamin A 6 . Previous case-control studies showed significantly increased risk of both chronic obstructive pulmonary disease (COPD) 41 and nontuberculous mycobacterial pulmonary disease 42 for persons with lower plasma vitamin A, and a cross-sectional study of 50 COPD patients and controls reported essentially the same 43 . Whether vitamin A inhibition of inflammatory responses and suppression of host respiratory infections can account for the respiratory mortality associations will require further study. The present finding of a weak and non-significant retinol-cancer mortality association after excluding the first 5 years of follow-up, indicating some potential reverse causality, is consistent with the lack of associations reported in previous studies [5][6][7]44 .
The major strength of our study is its basis in a large prospective cohort with complete follow-up up to 31 years and identification of cause-specific mortality via linkage with Finnish national registries. Retinol serological status was determined independently through a high-quality isocratic high-performance liquid chromatography platform for more than 29,000 men, which provided an accurate and objective measurement of retinol biochemical concentrations. Additionally, the large sample size enabled us to examine associations for the less common causes of mortality and test whether other factors such as age, alcohol consumption, or smoking intensity modified the vitamin A-mortality associations. Our findings are also subject to some limitations that should be noted. The cohort was a relatively homogenous population of male Finnish smokers which limits generalizability of our findings to other populations. There was only a single assessment of baseline serum retinol, and the physiological status of some participants may have changed during the long follow-up period. Other biomarkers relevant to vitamin A status, including retinol-binding protein, retinoic acid receptor expression, and C-reactive protein (CRP, a marker of inflammation, that the inflammation status may affect retinol homeostasis and serum retinol concentrations), that would have afforded a deeper evaluation of the mortality associations, were not available for the cohort. Underestimated mortality associations towards the null may result from biases attributed to nondifferential misclassification such as measurement error. However, we found the baseline and year-3 retinol concentrations were highly correlated (Spearman correlation coefficient = 0.69, P-value < 0.0001), indicating its relative stability during the postrandomization period. Despite having controlled for a large number of potential confounding factors, the possibility of findings being subject to residual confounding from other unmeasured factors cannot be ruled out.
In conclusion, our data indicate that higher vitamin A status is associated with lower overall and cause-specific mortality in men. The associations were based on over 30 years of follow-up and remained unchanged after adjusting for potential confounding risk factors and were largely stable across a range of population subgroups. The study provides evidence supporting the importance of vitamin A biochemical sufficiency for long-term health and longevity. Reexamination of these associations in populations that include women, nonsmokers, and other races/ethnicities is warranted.

Methods
Study cohort. The present analysis is nested within the Alpha-Tocopherol, Beta-Carotene Cancer Prevention (ATBC) Study cohort, details of which have been documented (ClinicalTrials.gov Identifier: NCT00342992) 45 . This randomized, double-blind, placebo-controlled trial evaluated whether alpha-tocopherol or betacarotene could decrease cancer incidence (particularly lung cancer). The study enrolled 29,133 male smokers, aged 50-69 years from 1985 to 1988 in southwestern Finland. The participants were assigned to one of four supplementation groupsalpha-tocopherol (50 mg/day), beta-carotene (20 mg/day), both vitamins, or placebo-for 5-8 years until the intervention ended on April 30, 1993. At the baseline visit, questionnaires obtained data for demographics, medical and smoking histories, and lifestyle characteristics, including a modified food frequency questionnaire and color picture booklet that queried consumption frequency and portion sizes for 203 food items and 73 mixed dishes for the previous 12 months. Height, weight, and blood pressure were measured by professional nurses, and presupplementation fasting blood samples were processed to serum and stored at −70°C. The ATBC Study was conducted in accordance with the requirement by the Declaration of Helsinki. All participants provided written informed consent, and the study was reviewed and approved by the Institutional Review Boards at the US National Cancer Institute and the Finnish National Public Health Institute.
Laboratory measurements. Serum concentrations of retinol, alpha-tocopherol, and beta-carotene were measured for 29,104 of the participants (99.9%) using reversed-phase high-performance liquid chromatography, and the coefficient of variation for serum retinol was 3.1% 45,46 . Statistical analysis. Cox proportional hazard regression models with attained age as the time metric were used to estimate hazard ratios (HRs) and two-sided 95% confidence intervals (CIs) to examine the association between serum retinol (quintiles and per one SD) and mortality. The proportional hazard assumption was not violated and was tested using model interaction terms of the dichotomized indicator of follow-up years and serum retinol quintiles, with a likelihood ratio test comparing models with and without the interaction terms. A set of a priori covariables were chosen for the risk models based on the existing knowledge and research regarding potential confounding of the serum retinol-mortality risk associations. In the multivariable analysis, we adjusted for age at baseline (continuous), body mass index (BMI, continuous), serum total and serum high-density lipoprotein (HDL) cholesterol (continuous), cigarettes smoked per day (continuous), years of smoking (continuous), alcohol intake (categorical), intervention assignment (alpha-tocopherol supplementation: yes or no; beta-carotene supplementation: yes or no), systolic and diastolic blood pressure (continuous), history of CVD (yes or no), and history of diabetes (yes or no). For each HR from the primary Cox regression model, the adjusted absolute risk difference (ARD) was calculated for each one SD increment of serum retinol for observation periods of 10, 20, and 30 years, with the 95% CIs for ARD being computed based on 500 bootstrap samples. In terms of missing values for any individual covariate (<5%), we generated a missing value indicator in the models.
Restricted cubic splines with four knots were applied to flexibly model the possible nonlinear association between serum retinol levels (continuous) and overall and cause-specific mortality, with the selected change-points being at the 5th, 25th, 75th, 95th percentiles of retinol concentration. Likelihood ratio tests compared differences between the linear term model and the model with both linear and cubic spline terms, with a P-value of < 0.0056 indicating nonlinearity [nine tests]. We further used multivariable fractional polynomials to model the potential nonlinear association between serum retinol and overall and causespecific mortality (by using R [version 3.5.2] Package "mfp"). Kaplan−Meier survival plots and log-rank tests assessed survival differences across serum retinol quintiles both for all participants and (separately) men without a baseline history of CVD or diabetes mellitus.
We also performed a mediation analysis to quantify what portion of the association between overall mortality and known risk factors (including obesity, intensity of smoking, duration of smoking, diabetes mellitus, alcohol drinking and physical activity) was mediated through serum retinol concentration. Using the above-mentioned Cox proportional regression model, we estimated the HRs and 95% CIs, and adjusted for age, obesity (no: BMI < 30; yes: BMI ≥ 30), BMI (continuous), years of smoking (quintile), cigarettes per day (quintile), history of diabetes (no or yes, self-reported at baseline), alcohol consumption (quintile), physical activity (light or moderate physical activity: no or yes), serum total and HDL cholesterol (continuous), intervention assignment, systolic and diastolic blood pressure (continuous), and history of CVD (no or yes, self-reported at baseline). The traditional risk factors examined were mutually adjusted. We used the %MEDIATE SAS macro to estimate the mediation proportion and 95% CIs of serum retinol for the traditional risk factor-mortality associations 47 .
All statistical tests were performed using SAS software (version 9.4; SAS Institute Inc, Cary, NC) and the R statistical language (version 3.5.2; Vienna, Austria). All reported P-values are two-sided at type I error rate of 0.05. The Bonferroni correction threshold was applied to control for multiple tests (0.05/ 9 = 0.0056 for primary and secondary analyses [nine tests], and 0.05/10 = 0.005 for ten effect modification tests in subgroups).
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
Because of previously enacted EU Data Protection Regulation (GDPR) privacy rules and an existing data use agreement between Finland and the U.S. National Cancer Institute, the ATBC Study data, and materials described in the manuscript may not be made indiscriminately publicly available for the purposes of reproducing the findings. Please contact the ATBC Study Principal Investigators for appropriate specific data requests (https://atbcstudy.cancer.gov/). Data requests are subject to approval by the study review committees. For example, the review committee will check if the data requests containing information that could compromise research participant privacy or participant informed consent. The response time to data requests may be affected during the COVID-19 pandemic.

Code availability
The analytical methods for the current study are available from the corresponding author upon appropriate request. Data analysis packages using R and SAS software are listed in the Reporting summary.