Associations between patterns of modifiable risk factors in mid-life to late life and longevity: 36 year prospective cohort study

Study population

The Nurses' Health Study began enrolling participants in 1976, when 121 700 female registered nurses, aged 30-55 years, residing in 11 US states were recruited to complete a baseline questionnaire with information on lifestyle and medical history. The Health Professionals Follow-up Study recruited 51 529 male healthcare professionals in 1986 (dentists, pharmacists, veterinarians, optometrists, osteopathic physicians, and podiatrists), aged 40-75 years at baseline. All participants returned a self-administered baseline questionnaire with a detailed medical history, lifestyle factors, and usual diet. In both cohorts, questionnaire data were collected at baseline and then every two years, to update information on modifiable risk factors and follow-up on the occurrence of chronic diseases. In this analysis, we used modifiable risk factor data collected from the 1984 and 1986 cycles as baseline for the Nurses' Health Study and Health Professionals Follow-up Study, respectively. We excluded participants who reported cardiovascular disease, cancer, or type 2 diabetes at baseline, and participants with extreme and implausible daily energy intakes (3500 kcal for women and 4200 kcal for men; 1 kcal=4.18 kJ).

Assessment of modifiable risk factors

Modifiable risk factors measured were body mass index, smoking status, alcohol intake, quality of diet, and physical activity. Participants were asked to self-report body weight and number of cigarettes currently smoked in questionnaires completed every two years. Physical activity was assessed by previously validated questionnaires to self-report the amount of time spent each week on different physical activities: walking, jogging, running, bicycling, callisthenics, aerobics, aerobic dance, use of rowing machine, lap swimming, playing tennis, and playing squash or racquet ball.9 Weekly energy expenditure in metabolic equivalent task hours (MET hours/week) was calculated.10 An 131 item food frequency questionnaire was given every four years to update dietary information from participants in the Nurses' Health Study and the Health Professionals Follow-up Study. Participants were asked how often (from "never or less than once per month" to "six or more times per day"), on average, they consumed a standard portion size of each food item during the previous year. Questions about consumption of alcoholic beverages (beer, wine, and liquor) were included in each questionnaire.

We used the Alternative Healthy Eating Index 2010 to measure the quality of the diet, based on intake levels of 10 components: fruit, vegetables, whole grains, long chain omega 3 fats, nuts and legumes, polyunsaturated fatty acids, sugar sweetened beverages, red and processed meat, trans fat, and sodium. The total score of the Alternative Healthy Eating Index 2010 ranged from 0 to 100, with a higher score indicating a better quality of diet.11 Online supplemental file 1 details information on how the modifiable risk factors were assessed. Online supplemental table S1 shows the time of the assessments of the modifiable risk factors in the Nurses’ Health Study and the Health Professionals Follow-up Study.

Assessment of mortality and longevity

Our primary endpoint was death from any cause by 31 October 2020 for the Nurses' Health Study and the Health Professionals Follow-up Study. In both cohorts, mortality data were collected from a systematic search of state vital records and the national death index database. The search was supplemented by reports from family members and postal authorities. These methods provided more than 98% of deaths in each cohort.12 A validation study conducted by physicians who were blinded to data on risk factors reviewed the death certificates and medical records to classify the cause of death according to ICD-8 and ICD-9 (international classification of diseases, eighth and ninth revisions).

Longevity was defined as survival to age ≥85 years, according to the life expectancy of the US population and the vital status of our data.1 We found that 66% of participants in the Nurses' Health Study and 68% of participants in the Health Professionals Follow-up Study had achieved longevity (ie, participants were alive aged >85 years by 31 October 2020), and we excluded participants whose longevity status was not known at the end of follow-up. We also looked at longevity in participants who did not have a diagnosis of type 2 diabetes (ICD-8 code 250, ICD-9 code 2500), cardiovascular disease (ICD-8 codes 390-459 or 795, ICD-9 codes 3900-4590 or 7950), or cancer (ICD-8 codes 140-207, ICD-9 codes 1400-2070).

Assessment of covariates

In the Nurses' Health Study, age, race, use of aspirin, menopausal status, use of postmenopausal hormones, and family history of myocardial infarction were collected in the baseline questionnaire in 1984. Information on annual family income, use of multivitamins, and a family history of cancer or type 2 diabetes were collected in 1986. Information on education was asked in the questionnaire in 1992. In the Health Professionals Follow-up Study, age, race, work status, family history of myocardial infarction, cancer, or type 2 diabetes, and use of aspirin, antihypertensive agents (including β blockers, calcium channel blockers, and nitrates), and cholesterol lowering drugs were collected with the questionnaire at baseline in 1986, and use of multivitamins use was collected in 1988.

Statistical analysis

We identified patterns of risk factors with smooth mixed models, which describes trajectories with high flexibility with the use of smoothing functions of time and thus improves the accuracy of group classification.8 With the smoothing mixture model, we assumed smoothing functions for age and allowed for random effects of individuals classified in the same group. The R script of the smoothing mixture model was accessed in Github (https://github.com/mingding-hsph/Smoothing-mixture-model). We investigated the sources of variances for all risk factors, and found that the between-person variance was significantly higher than within-person variance for each risk factor We investigated the sources of variances for all risk factors, and found that the between-person variance was significantly higher than within-person variance for each risk factor (online supplemental table S2), indicating that patterns of risk factors identified would be largely driven by variations between individuals. Hence to capture patterns of a risk factor over time most efficiently, we simultaneously identified trajectories of the risk factor and trajectories of change in the risk factor (calculated as the difference in risk factor from baseline at each assessment during follow-up) with the smoothing mixture model, and used the joint group memberships to classify participants.

To allow for meaningful interpretation, for each risk factor, we assumed three trajectories for the risk factor (patterns with high, medium, and low values) and three trajectories of change in the risk factor (patterns with increase, no change, and decrease in the factor from baseline), which gave nine joint patterns: high-stable, high-increase, high-decrease, medium-stable, medium-increase, medium-decrease, low-stable, low-increase, and low-decrease. Because the smoothing mixture model is a new method, we also identified trajectories of risk factors and trajectories of change in risk factors with group based trajectory analysis.13 The online supplemental file provides detailed information on the smoothing mixture model, variance decomposition, joint group membership of trajectories of risk factors and trajectories of change in risk factors, and group based trajectory analysis.

To minimise the possibility of reverse causation, we censored risk factors reported after a diagnosis of cardiovascular disease, type 2 diabetes, or cancer, and risk factors reported after age 85 years. We also excluded individuals with less than two measurements for derivation of trajectories (online supplemental figures S1, S2). To remove the effects of temporal trends in dietary intake and physical activity,14 we standardised physical activity and dietary factors (alcohol intake and Alternative Healthy Eating Index score) during each follow-up cycle,15 and mapped the standardised values to real values based on the mean and standard deviation over the whole follow-up period with the inverse cumulative distribution function.

For each risk factor, we examined associations between patterns of risk factor and longevity with logistic regression models, and associations between patterns of risk factor and healthy longevity (defined as individuals who did not have a diagnosis of cardiovascular disease, type 2 diabetes, or cancer) with multinomial logistic models. The models were adjusted for age (continuous), race (white v non-white), family history of cancer (yes, no), myocardial infarction (yes, no), or type 2 diabetes (yes, no), multivitamin use (yes, no), menopausal status (yes, no, women only), postmenopausal hormone use (yes, no, women only), education (registered nurse, bachelor degree, master degree and higher, women only), social economic status (annual family income (four groups) for women and work status (disabled, retired, part time, full time) for men), and the other four factors at baseline (continuous variables).

In models that assessed the association between patterns of change in risk factor and longevity, we also adjusted for the pattern of that factor (categorical). We examined associations between patterns of risk factor and life expectancy with the accelerated failure time model. We modelled survival time as the number of years lived from baseline until death or to the end of follow-up, whichever came first. We assumed an exponential distribution of survival time and applied transformation 100×(e^β−1) to the regression coefficient β, which can be interpreted as the per cent increase in life expectancy comparing two groups. We pooled the data of both cohorts to obtain overall effects. All statistical tests were two sided at a type I error rate of 0.05. The smoothing mixture model was performed with R 3.5.0, and logistic regression and accelerated failure time model were applied with SAS version 9.2 for UNIX (SAS Institute, Cary, NC).

Patient and public involvement

No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community.

Principal findings

Our analysis followed 85 346 men and women over a period of 36 years, and we found that 46 042 participants achieved longevity (alive at 85 years) and 25 322 participants achieved healthy longevity (ie, no diagnosis of cardiovascular disease, type 2 diabetes, or cancer). During mid-life and late life, we found that maximum longevity was achieved in participants maintaining a normal body mass index, who never smoked, ate a healthy diet, and had physical activity levels and alcohol consumption that met public health recommendations. The similarity of the associations between risk factors and longevity for both women and men further strengthens our findings. Previous studies have shown that modifiable risk factors, including not smoking and being physically active, were associated with longer survival.18 19 Our study adds new evidence to the topic from the perspective of the course of a lifetime.

We found that participants with a low-stable pattern for body mass index were most likely to achieve longevity and healthy longevity, and this finding is consistent with previous studies involving millions of participants that showed that body mass index in the range 21-25 was associated with the lowest risk of mortality.20–24 Moreover, our study suggested no further gain in longevity for the low-increase, medium-increase, and high-increase patterns in body mass index, in agreement with previous studies showing that even moderate weight gain during mid-life was associated with a much higher risk of mortality and major chronic diseases.25 Our findings for participants with low-decrease, medium-decrease, and high-decrease patterns indicated that participants with weight loss were less likely to achieve longevity. The reason might be reverse causation, because participants with underlying medical conditions might lose weight before the symptoms of disease appeared. Although we censored body mass index before a diagnosis of cardiovascular disease, cancer, or type 2 diabetes, reverse causation bias is a plausible explanation, and previous research showed that unintentional weight loss was associated with higher mortality rates.26

Comparison with other studies

The association between physical activity with mortality has been extensively studied,27 and the 2008 physical activity guidelines for Americans recommends a minimum of 7.5 MET hours/week of physical activity for health benefits (equivalent to 150 minutes/week of moderate intensity activity or 75 minutes/week of vigorous physical activity), and 15.0 MET hours/week for more benefits.17 However, whether greater health benefits can be achieved with physical activity levels >15.0 MET hours/week or whether a threshold of physical activity exists for mortality benefits is uncertain. Previous studies showed a non-linear association, with the lowest risk of mortality at 21-35 MET hours/week.28 29 We also found that a medium-increase pattern (20~30 MET hours/week) in mid-life to late life was associated with the highest probability of longevity and healthy longevity, and that the benefit to longevity persisted for the high patterns, although no further gains in longevity were found. Moreover, our study showed that an increase in physical activity during a lifetime was associated with greater longevity, but the benefits to longevity in mid-life were lost for participants who reduced their physical activity levels. Our findings are consistent with the EPIC (European Prospective Investigation into Cancer) Norfolk study,30 which looked at three measures of physical activity and assumed linear associations for participants' trajectories. One explanation for reduced life expectancy in those with a decrease pattern could be reverse causation, because participants with chronic diseases, such as cardiovascular disease and cancer, might reduce their amount of physical activity. Another explanation is that recent physical activity might have a more important role in affecting health.

We detected a dose-response relation between patterns for quality of diet and longevity, with those with a high-increase pattern achieving highest longevity and healthy longevity. Our findings are in line with previous studies showing that a higher quality diet was associated with a lower risk of mortality31 32 and that improving the quality of your diet was associated with a lower risk of death.6 Randomised clinical trials have shown that a healthy diet promotes weight loss,33 substantially lowers blood pressure,34 35 and efficiently prevents cardiovascular disease.36 A plausible mechanism is that the beneficial effects might be mediated through improvement in endothelial function, inflammatory markers, and insulin resistance,37 and the foods and nutrients promoted by the Alternative Healthy Eating Index (fruits, vegetables, whole grain, and unsaturated fatty acids) might account for the effect. Overall, our study adds to the evidence that individuals should adhere to a healthy diet in mid-life to late life to boost longevity.

The health benefits of alcohol intake are controversial, particularly for moderate intake of alcohol. Our study showed that individuals with the medium pattern were more likely to achieve longevity and healthy longevity than those with the low and high patterns, supporting the 2020-25 dietary guidelines for Americans of no more than two drinks (~30 g) a day for men and no more than one drink (~15 g) a day for women.16 Our findings are in agreement with previous studies. One meta-analysis involving more than a million participants showed a non-linear association between intake of alcohol and total mortality, with the lowest risk at 1-2 drinks/day for women and 2-4 drinks/day for men.38 Prospective studies found that moderate intake of alcohol was associated with the lowest risk of total mortality and the highest probability of achieving longevity.39 40 In the Global Burden of Disease study, however, with data pooled from 195 countries, the risk of mortality increased with higher levels of consumption, including moderate intake of alcohol.41 The reason for the inconsistency is because of the varied distribution of causes for cause specific mortality between developing and developed countries. In fact, the Global Burden of Disease study found that moderate intake of alcohol was associated with a lower risk of cardiovascular disease and type 2 diabetes, the main drivers of mortality in developed countries.41 Longevity of moderate intake of alcohol can have benefits; alcohol raises levels of high density lipoprotein and improves insulin sensitivity, which protect against cardiovascular disease and type 2 diabetes.42 Alcohol intake with family and friends has social and psychological benefits that might also contribute to health and wellbeing.43 We found that high patterns of alcohol intake were less likely to be associated with longevity, consistent with previous studies,38–40 and evidence has been convincing that alcohol consumption increases the risk of breast cancer.44

Smoking was the leading modifiable risk factor for mortality worldwide in 2015,45 and 11.5% of global deaths were attributed to smoking.46 In our study, smoking showed the strongest association with longevity compared with other risk factors, highlighting the importance of stopping smoking in the promotion of longevity. We found that an increase in intensity of smoking shortens the lifespan and a decrease in smoking extends life. Although the prevalence of smoking has been decreasing over the past 30 years, the rate is still high (ie, 25.0% for men and 5.4% for women worldwide).46 Our study supports an important public health message that it is never too late to stop smoking and gain benefits for healthy longevity.

Strengths and limitations of the study

Our study benefitted from appropriate use of advanced statistical analyses and strict sensitivity analyses. Firstly, we creatively captured individual’s risk factor patterns by classifying participants based on joint group membership of trajectories of risk factors and change in risk factors. We used the smoothing mixture model to derive patterns of risk factors, which models trajectories with high flexibility and reduces the probability of misclassification of trajectories.8 Secondly, although participants who achieved longevity tended to have more repeated measures of risk factors than those who died before age 85 years, classification of risk factor patterns was independent of the number of measurements and thus independent of outcome.8 Thirdly, because censoring time is dependent on longevity as an outcome, implementing survival analysis or competing risk analysis to examine longevity can result in selection bias.47 Thus the use of logistic regression and multinomial regression dealt with the possibility of biased estimates. To leverage censoring time, however, we also used the accelerated failure time model to examine associations between risk factor patterns and life expectancy. Fourthly, because the start age of each individual’s trajectory varied, we examined whether age at baseline affected the associations between joint patterns of risk factors and longevity, and found that participants should adhere to a healthy lifestyle at an earlier age to increase the odds of longevity.

Our study had several limitations. Firstly, the Nurses' Health Study and Health Professionals Follow-up Study included predominantly white healthcare professionals, which could limit the generalisability of our findings to other nationalities and races or ethnicities. However, participants’ health related occupations were an advantage that allowed us to collect high quality data with self-reported questionnaires and enhance the internal validity of the study by reducing confounding. Secondly, risk factors were self-reported by questionnaire, and measurement error was inevitable. Our food frequency questionnaires have been extensively validated against diet records,48–51 however, and body mass index and physical activity have been validated against standard measurements.9 52 Thirdly, given the observational design of the study, we could not directly establish causal relations between risk factor patterns and longevity. Fourthly, we censored risk factors after a diagnosis of cardiovascular disease, type 2 diabetes, or cancer to avoid reverse causation, but this method might result in missing data. We imputed missing data and found that the shape of joint patterns of risk factors and the associations with longevity were similar to our main findings, indicating the robustness of our main findings. Fifthly, we excluded participants whose longevity was unknown, which might cause selection bias. The main reason for the exclusion was that these participants were younger at baseline and were alive without reaching the age of 85 years at the end of follow-up. The population characteristics of the excluded participants and the main study population were similar, however, and we believe that the excluded population would likely give similar findings to our main population. Our sensitivity analyses (online supplemental tables S14 and S15) suggest that exclusion of these participants had minimal effects on our findings.

Conclusion

We found that maximum longevity was achieved in participants maintaining a normal body mass index, who never smoked, ate a healthy diet, and had physical activity levels and alcohol intake that met public health recommendations through mid-life and late life.