Secondary Logo

Journal Logo


Hypothetical Midlife Interventions in Women and Risk of Type 2 Diabetes

Danaei, Goodarza,b; Pan, Anc; Hu, Frank B.b,c,d; Hernán, Miguel A.b,e,f

Author Information
doi: 10.1097/EDE.0b013e318276c98a


Diabetes is a major cause of death and disability worldwide,1,2 and its prevalence has increased substantially in most regions of the world in the last three decades.3 Complications of diabetes put a major economic burden on health systems in both developed and developing countries.4–6 Primary prevention of type 2 diabetes, which constitutes >90% of cases, is a major concern for health systems worldwide.

Several randomized trials have examined the effect of physical activity, smoking cessation, and healthy diet on the incidence of type 2 diabetes in high-risk participants over relatively short periods of about 3 years.7–15 Overall, these randomized trials reported about 50% reduction in the incidence of type 2 diabetes after intensive lifestyle modifications.16–18 Prospective observational studies have mostly examined the long-term association between lifestyle and incidence of type 2 diabetes19–22 in relatively healthy populations, but these studies did not specify the corresponding interventions or the time of their initiation.

Devising an informed policy for diabetes prevention requires estimating the effect of lifestyle interventions initiated in midlife or later (as randomized trials did) over long periods and in relatively healthy populations (as observational studies did). We applied the parametric g-formula to the data obtained from the observational Nurses’ Health Study to estimate the 24-year risk of type 2 diabetes under various hypothetical lifestyle interventions that start in midlife or later.


Study Population

The Nurses’ Health Study is a prospective cohort study that started in 1976 by enrolling 121,700 U.S. female nurses age 30–55 years. A questionnaire was mailed to collect data on sociodemographic, lifestyle, and dietary factors and on history of diseases and treatments. Biennial questionnaires have been mailed since then to update information on risk factors and disease incidence. More details on this cohort are available elsewhere.23

We used the 1984 questionnaire as baseline, because a detailed 131-item food frequency questionnaire (FFQ) was distributed in that year. Women were excluded from our analysis if they had a diagnosis of diabetes before 1984 or did not have information on diabetes diagnosis date. In addition, we excluded participants who left >70 items blank on the baseline FFQ, or who reported unusual total energy intakes (ie, energy intake <500 or >3500 kcal/day). We also excluded participants without baseline information on date of birth, body weight and height, smoking status, physical activity, and dietary variables, and those who had cancer or cardiovascular disease at baseline (see eFigure 1, for a flowchart of participant selection). After these exclusions, 76,402 women were available for the analysis. Women were followed until the occurrence of type 2 diabetes, death, incomplete follow-up (ie, not returning a questionnaire), or administrative end of follow-up in June 2008, whichever happened first.

Diet was measured using a semiquantitative FFQ in 1984, 1986, and every 4 years afterward. The FFQ asked about the usual intake of various food items, including alcoholic drinks, during the past 12 months. The reproducibility and validity of the questionnaire have been shown elsewhere.24 Dietary data recorded in the 1980 questionnaire were used to adjust for pre-baseline diet. Physical activity was reported in 1980, 1982, 1986, 1988, and every 4 years thereafter, using a validated questionnaire25 on type, frequency, and intensity of each activity. We summed the duration of moderate or vigorous activities per week (ie, requiring at least three Metabolic Equivalent of Task scores per hour, including brisk walking). Body weight was self-reported in each biennial questionnaire and height was self-reported in 1976.

We truncated the values of dietary risk factors, body mass index (BMI), and physical activity in each period at the 99th percentile to prevent implausible values from affecting our analyses. Sensitivity analyses using various thresholds based on expert knowledge, or setting values above the 99th percentile to missing and carrying the last observed value forward, did not change the results materially.

Diabetes Ascertainment

Details of diabetes ascertainment has been described elsewhere.26 Briefly, a supplementary questionnaire was mailed to participants who reported a diagnosis of diabetes. A case of type 2 diabetes was considered confirmed if, according to the National Diabetes Data Group criteria,27 at least one of the following was reported on the supplementary questionnaire: (1) one or more classic symptoms plus fasting plasma glucose levels of ≥7.8 mmol/l or random plasma glucose levels of ≥11.1 mmol/l; (2) at least two elevated plasma glucose concentrations on different occasions (fasting plasma glucose ≥7.8 mmol/l, random plasma glucose levels of ≥11.1 mmol/l, or concentrations of ≥11.1 mmol/l after 2 hours or more shown by oral glucose tolerance testing) in the absence of symptoms; or (3) treatment with insulin or oral hypoglycemic agents. The diagnostic criteria changed in June 1998: according to the American Diabetes Association criteria,28 a fasting plasma glucose of 7.0 mmol/l instead of 7.8 mmol/l was considered the threshold for the diagnosis of diabetes. Only confirmed cases were included in the analysis. We excluded cases designated as gestational or secondary diabetes. The validity of the supplementary questionnaire has been previously documented by reviewing medical records and assessing undiagnosed diabetes in a random sample of women.26,27,29 Deaths were identified by reports from next of kin or postal authorities, or by searching the National Death Index. At least 98% of deaths among the study participants were identified.30

Hypothetical Lifestyle Interventions

We considered eight hypothetical interventions and their combinations, based on the evidence from both randomized trials and observational studies on their potential effect for diabetes prevention19,20,31–35: quitting smoking, losing 5% of BMI every 2 years if overweight or obese (defined as BMI ≥25 kg/m2), exercising at least 30 minutes a day (moderate or vigorous), eating at least two servings of whole grain per day, drinking at least two cups of coffee per day, drinking at least 5 grams of alcohol per day, eating at most three servings of red meat per week (including unprocessed and processed), and drinking at most one serving of soda per week. All interventions started at baseline in 1984 and continued until the end of follow-up. Except for the weight-loss intervention, which imposed a gradual decline in body weight, all other interventions imposed a minimum or maximum threshold on the level of a risk factor. For these interventions, values beyond the threshold were set to the threshold level.

The intensities of the interventions were chosen to match current public health guidelines (eg, Centers for Disease Control and Prevention guidelines for physical activity, World Health Organization definition of overweight and obesity) or were selected to reflect feasible public health interventions. We estimated the effects of more intensive interventions for weight loss and physical activity in sensitivity analyses.

Statistical Analysis

We used the parametric g-formula,36 a generalization of standardization for time-varying exposures and confounders, to estimate the 24-year risks of type 2 diabetes under the selected lifestyle interventions. The parametric g-formula has been previously used to estimate the effect of multiple lifestyle interventions on the risk of coronary heart disease.37 If all time-varying confounders have been correctly measured and modeled at all time-points, the g-formula can be used to consistently estimate the standardized risk of type 2 diabetes under hypothetical interventions. We used a Kaplan-Meier estimator that incorporates censoring by death and loss to follow-up.38 For the exact formula see electronic appendix eText (

A simplified description of the process of estimating risks using parametric g-formula is as follows: we start by fitting regression models for all potential confounders and for the disease outcome using data on the entire study population. We then use these models to simulate the risk of the disease under various interventions in five steps: (1) take the observed joint distribution of covariates at baseline; (2) estimate the joint distribution of time-varying covariates at the next time-point using the parametric models; (3) “intervene” by setting the values of some covariates to values determined by the hypothetical interventions; (4) estimate the predicted probability of the outcome using these new values; (5) repeat steps 2 through 4 for the entire duration of follow-up to estimate the predicted risk of the disease under the selected interventions. See articles by Taubman et al37 and Young et al39 for a more detailed description.

More formally, the standardized cumulative risk estimated by the g-formula is a weighted average of the risks of type 2 diabetes conditional on the specified intervention values and the observed confounder history. The weights are the probability density functions of the time-varying confounders, which are estimated via parametric regression models. We approximated the weighted average by using a Monte Carlo simulation of 10,000 individuals with the baseline values of covariates sampled from their empirical distribution. The values of time-varying covariates for each 2-year interval were drawn from the distribution estimated via the regression models after setting the values of lifestyle factors to those specified by the interventions.

The models included calendar year for each period and the following potential baseline confounders: age, history of diabetes in first-degree relatives, smoking and oral contraceptive use before 1980, marital status, education, husband’s education, employment, and stress in daily life and work, as well as pre-baseline values of the variables corresponding to the eight selected intervention (ie, BMI, smoking, exercise, and intake of meat, whole grain, coffee, alcohol, and soda). These latter adjustments enabled us to estimate the effect of changes in lifestyle in middle-aged women as opposed to the lifelong effect of a healthy lifestyle.

We modeled the distribution of 20 time-varying covariates: multivitamin use, aspirin use, statin use, postmenopausal hormone use, smoking, physical activity, soda intake, coffee intake, red meat intake, whole-grain intake, alcohol use, BMI, high blood pressure, high serum cholesterol, coronary heart disease, stroke, angina or coronary artery bypass graft, cancer, menopause, and osteoporosis (see eTable 1,

We compared the estimated risks under various interventions with the risk under no intervention to calculate the population risk ratio and population risk difference. The population attributable risk is one minus the population risk ratio. To estimate the 95% confidence intervals (CIs), we used nonparametric bootstrapping with 500 samples.40 We also computed the proportion of women who were hypothetically intervened on in any period and the average proportion of women intervened on in each 2-year period. The latter measures overall adherence in the observed data to the hypothetical intervention among those following the intervention up until the previous period. We examined the possibility of effect modification by conducting the analysis separately in subsets of the study population defined at baseline according to age, BMI, and family history of diabetes. All analyses were conducted using SAS 9.2 (Cary, NC). The SAS macro and its documentation are available at


Table 1 shows the baseline characteristics of the 76,402 women who met all eligibility criteria in 1984. Mean age was 49.9 years and mean BMI was 24.4 kg/m2; 25% of women were current smokers and 19% had a family history of diabetes. During the 24 years (1.4 million person-years) of follow-up, there were 6044 cases of type 2 diabetes and 8260 deaths. In total, 17,690 participants did not reach the end of follow-up (ie, did not return a questionnaire). The models closely replicated the mean of the risk factors under no intervention. For example, the mean difference between the observed and simulated number of cigarettes smoked per day was <0.4 cigarettes during follow-up; similarly, the point estimate of the ratio of the observed to simulated mean BMI was never smaller than 0.99 or larger than 1.00 (see eFigures 2–9, The simulated 24-year risk of diabetes was the same as the observed risk at 9.6%. The coefficients of the models used in the simulations are presented in eTable 2 (

Baseline Characteristics of the Eligible 76,402 Women, Nurses’ Health Study 1984

Table 2 shows the 24-year risk of diabetes under various hypothetical lifestyle interventions. Among the nondietary interventions, weight loss was estimated to reduce the risk by 24% (95% CI = 22 to 26%), exercise by 19% (6 to 30%), and quitting smoking by 0% (–1 to 2%) when compared with no intervention. Among the selected dietary interventions, drinking at least 5 grams of alcohol a day was estimated to reduce the risk by 19% (12 to 23%), eating less than three servings of red meat per week by 8% (5 to 11%), and drinking at least two cups of coffee a day by 3% (0 to 6%). The mean alcohol intake under the intervention “drink at least 5 grams of alcohol a day” ranged from 9.2 to 11.7 grams a day during follow-up, which is equivalent to one drink a day.

Diabetes Risk Under Hypothetical Lifestyle Interventions, Nurses’ Health Study, 1984–2008

We estimated that the 24-year risk of type 2 diabetes would be reduced by 39% (29 to 47%) under the three nondietary interventions, by 29% (21 to 35%) under the five dietary interventions, and by 55% (47 to 63%) under all eight interventions. The estimated 24-year risk of type 2 diabetes under all interventions was 4.3% (3.6 to 5.1%). Of all participants, 25% maintained a BMI of 25 kg/m2, fewer than 11% followed each of the dietary interventions, and 0% followed all eight interventions for the whole duration of follow-up (9% followed all eight interventions at some point during follow-up).

Table 3 presents the results of the analyses for more extreme weight-loss interventions. We estimated that reducing BMI at 5% per 2-year period down to 23 kg/m2 would reduce the risk of diabetes by 53% (51 to 56%), whereas reducing BMI at a faster pace of 10% every 2 years would reduce the risk by 60% (57 to 63%). Combining this latter intervention with the seven other lifestyle interventions would reduce the risk of type 2 diabetes by 72%. We estimated a risk reduction of 16% (8 to 22%) under a hypothetical intervention in which participants exercise for 30 minutes per day if they have a normal BMI (ie, BMI <25 kg/m2) and 1 hour per day if they are overweight or obese (ie, BMI ≥25 kg/m2). A more intensive intervention to engage in at least 1 hour per day of moderate or vigorous activity, regardless of BMI, would reduce the risk of type 2 diabetes by 15% (7 to 22%).

Effect of Hypothetical Weight-Loss Interventions on Diabetes Incidence in the Nurses’ Health Study 1984–2008

To evaluate the sensitivity of the estimates to our analytic decisions, we conducted analyses that varied the arbitrary ordering of the variables measured in each questionnaire, estimated censoring due to incomplete follow-up, considered a different intervention on alcohol where women would drink 5–10 grams of alcohol per day, and included women who did not return one or two questionnaires by carrying their last reported values forward. The estimates of relative risks and risk differences did not change materially.

The effects of lifestyle interventions were stronger in younger women (<50 years old) and those who were overweight or obese at baseline, in both the risk ratio and risk difference scales (Table 4). For all eight interventions combined, the 24-year risk of type 2 diabetes would be reduced by 10.8 percentage points in women who were overweight or obese as compared with only 1.9 percentage points in women who had a BMI of <25 kg/m2 at baseline (Table 4). Women with a family history of diabetes were also estimated to benefit more from these interventions: the risk difference for all eight interventions in this group was 7.3% versus 4.8% in those without a family history of diabetes.

Effect of Joint Hypothetical Interventions in Subgroups Defined at Baseline by (A) Body Mass Index, (B) Age, and (C) Family History Of Diabetes


Our results suggest that, in this cohort of U.S. women, 55% of cases of type 2 diabetes could have been prevented by a combination of dietary and nondietary lifestyle modifications. Our estimates are particularly relevant for health policy because they quantify the 24-year impact of lifestyle interventions that start in midlife or later in relatively healthy women.

A beneficial effect of lifestyle modification on diabetes risk had been previously found in several randomized trials. However, these studies were designed to evaluate a short-term effect (over approximately a 3-year period) in overweight participants with impaired glucose tolerance. Both the Diabetes Prevention Program7 and the Diabetes Prevention Study8 found that a combined intervention on diet and physical activity reduced diabetes risk by 58% when compared with general guidance or written advice. The risk of diabetes in the control group was 29% at 3 years in the Diabetes Prevention Program and 23% at 4 years in the Diabetes Prevention Study, which are much higher than our 9.6% risk at 24 years in a population of healthy U.S. women.

Our estimates may not be generalizable to other populations with different distributions of risk factors, as the g-formula standardizes the risk to the distribution of risk factors in the particular population under study. For example, we estimated no reduction in risk of type 2 diabetes if all women had quit smoking but only 25% of women in our population were smokers. When we compared the risk had everyone been “forced” to smoke 20 cigarettes a day to the risk had everyone quit smoking, the estimated population risk ratio was 1.1 (data not shown). Also, the magnitude of our estimates is specific to the set of interventions that was considered. Though our results are generally consistent with previous analyses of prospective studies,34,35,41,42 differences do exist because previous studies did not specify the time of initiation of the lifestyle interventions and considered more extreme weight loss.19–22 For example, a previous analysis of the Nurses’ Health Study cohort classified women as low-risk if they met five criteria (BMI <25 kg/m2, diet high in cereal fiber and polyunsaturated fat and low in trans-fat and glycemic load, at least half an hour per day of moderate-to-vigorous physical activity, no current smoking, at least a half-serving of an alcoholic beverage per day) between 1986 and 1996. As compared with the rest of the cohort, women in the low-risk group had an average 2-year relative risk of diabetes of 0.09 (95% CI = 0.05 to 0.17). The authors estimated that some hypothetical intervention (different from the ones we considered here) on the above five factors could have prevented 91% of diabetes cases.

Our analysis has several strengths. The Nurses’ Health Study cohort has collected detailed data on usual dietary intakes, physical activity, body weight, and smoking every 2–4 years, and diabetes diagnosis was validated. We included 24 years of follow-up and had a large number of cases, which allowed meaningful subgroup analyses. By applying the parametric g-formula,37,43 we could estimate the effect of hypothetical interventions starting in middle-age or later while (1) appropriately adjusting for time-varying confounders affected by prior exposure, (2) generating adjusted estimates of absolute risk and population attributable risk, and (3) estimating effects of interventions individually and in various combinations.

Like other analyses of observational data, the validity of our results relies on the key assumptions of no residual confounding, no measurement error, and no model misspecification. The possibility of residual confounding cannot be logically excluded despite adjustment for many potential confounders. A certain degree of measurement error is expected for lifestyle variables and may have contributed to bias. We had to rely on self-reported weight and height; a validation study on a small sample of the study population showed a 0.96 correlation between self-reported and measured weight.44 Though the prediction of our models under the hypothetical interventions cannot be directly evaluated (0% of women followed all interventions during the entire follow-up), our models provided accurate predictions under no intervention, a necessary condition for no model misspecification.

We tried to simplify the public health interpretation of our estimates by comparing interventions with well-defined start times and exposure values, by using food items rather than dietary scores19 or the Healthy Eating Index,45 and by including total caloric intake in the models so that our hypothetical interventions imply replacing the selected food item with other foods that are usually taken as a substitute in the study population. However, the interpretation of our estimates for weight loss and exercise is complicated because there are multiple versions of the interventions. For example, participants may lose weight by reducing their caloric intake or by using weight-loss medications; similarly, the selected duration and level of exercise may be achieved by increasing the frequency or duration of different types of activities.46

In summary, our results suggest that 55% of cases of diabetes that occurred during 24 years of follow-up in this large prospective study of U.S. women could have been prevented by eight lifestyle interventions initiated in midlife or later. The most effective intervention in this population was losing weight, followed by eating a healthy diet, and engaging in regular moderate or vigorous physical activity.


We are grateful to Roger Logan and Jessica Young for their technical assistance; Walter Willett and Meir Stampfer for their comments on a previous draft; the Channing Laboratory, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School; and all the women enrolled in the Nurses’ Health Study.


1. Lopez AD, Mathers CD, Ezzati M, Jamison DT, Murray CJ. Global and regional burden of disease and risk factors, 2001: systematic analysis of population health data. Lancet. 2006;367:1747–1757
2. Danaei G, Lawes CM, Vander Hoorn S, Murray CJ, Ezzati M. Global and regional mortality from ischaemic heart disease and stroke attributable to higher-than-optimum blood glucose concentration: comparative risk assessment. Lancet. 2006;368:1651–1659
3. Danaei G, Finucane MM, Lu Y, et al. National, regional, and global trends in fasting plasma glucose and diabetes prevalence since 1980: systematic analysis of health examination surveys and epidemiological studies with 370 country-years and 2.7 million participants. Lancet.. 2011;378:31–40
4. Abegunde DO, Mathers CD, Adam T, Ortegon M, Strong K. The burden and costs of chronic diseases in low-income and middle-income countries. Lancet. 2007;370:1929–1938
5. Bjork S, Kapur A, King H, Nair J, Ramachandran A. Global policy: aspects of diabetes in India. Health Policy. 2003;66:61–72
6. MacLeod KM, Tooke JE. Direct and indirect costs of cardiovascular and cerebrovascular complications of type II diabetes. Pharmacoeconomics. 1995;8(suppl 1):46–51
7. Knowler WC, Barrett-Connor E, Fowler SE, et al.. Diabetes Prevention Program Research Group. Reduction in the incidence of type 2 diabetes with lifestyle intervention or metformin. N Engl J Med. 2002;346:393–403
8. Tuomilehto J, Lindström J, Eriksson JG, et al.. Finnish Diabetes Prevention Study Group. Prevention of type 2 diabetes mellitus by changes in lifestyle among subjects with impaired glucose tolerance. N Engl J Med. 2001;344:1343–1350
9. Pan XR, Li GW, Hu YH, et al. Effects of diet and exercise in preventing NIDDM in people with impaired glucose tolerance. The Da Qing IGT and Diabetes Study. Diabetes Care. 1997;20:537–544
10. Dyson PA, Hammersley MS, Morris RJ, Holman RR, Turner RC. The Fasting Hyperglycaemia Study: II. Randomized controlled trial of reinforced healthy-living advice in subjects with increased but not diabetic fasting plasma glucose. Metab Clin Exp. 1997;46(12 suppl 1):50–55
11. Kosaka K, Noda M, Kuzuya T. Prevention of type 2 diabetes by lifestyle intervention: a Japanese trial in IGT males. Diabetes Res Clin Pract. 2005;67:152–162
12. Ramachandran A, Snehalatha C, Mary S, Mukesh B, Bhaskar AD, Vijay V. The Indian Diabetes Prevention Programme shows that lifestyle modification and metformin prevent type 2 diabetes in Asian Indian subjects with impaired glucose tolerance (IDPP-1). Diabetologia. 2006;49:289–297
13. Bo S, Ciccone G, Baldi C, et al. Effectiveness of a lifestyle intervention on metabolic syndrome. A randomized controlled trial. J Gen Intern Med. 2007;22:1695–1703
14. Roumen C, Corpeleijn E, Feskens EJ, Mensink M, Saris WH, Blaak EE. Impact of 3-year lifestyle intervention on postprandial glucose metabolism: the SLIM study. Diabet Med. 2008;25:597–605
15. Lindahl B, Nilssön TK, Borch-Johnsen K, et al. A randomized lifestyle intervention with 5-year follow-up in subjects with impaired glucose tolerance: pronounced short-term impact but long-term adherence problems. Scand J Public Health. 2009;37:434–442
16. Yamaoka K, Tango T. Efficacy of lifestyle education to prevent type 2 diabetes: a meta-analysis of randomized controlled trials. Diabetes Care. 2005;28:2780–2786
17. Orozco LJ, Buchleitner AM, Gimenez-Perez G, et al. Exercise or exercise and diet for preventing type 2 diabetes mellitus. Cochrane Database Syst Rev. 2008
18. Gillies CL, Abrams KR, Lambert PC, et al. Pharmacological and lifestyle interventions to prevent or delay type 2 diabetes in people with impaired glucose tolerance: systematic review and meta-analysis. BMJ. 2007;334:299
19. Hu FB, Manson JE, Stampfer MJ, et al. Diet, lifestyle, and the risk of type 2 diabetes mellitus in women. N Engl J Med. 2001;345:790–797
20. Mozaffarian D, Kamineni A, Carnethon M, Djoussé L, Mukamal KJ, Siscovick D. Lifestyle risk factors and new-onset diabetes mellitus in older adults: the cardiovascular health study. Arch Intern Med. 2009;169:798–807
21. Laaksonen MA, Knekt P, Rissanen H, et al. The relative importance of modifiable potential risk factors of type 2 diabetes: a meta-analysis of two cohorts. Eur J Epidemiol. 2010;25:115–124
22. Steinbrecher A, Morimoto Y, Heak S, et al. The preventable proportion of type 2 diabetes by ethnicity: the multiethnic cohort. Ann Epidemiol. 2011;21:526–535
23. Colditz GA, Manson JE, Hankinson SE. The Nurses’ Health Study: 20-year contribution to the understanding of health among women. J Womens Health. 1997;6:49–62
24. Willett WC, Sampson L, Stampfer MJ, et al. Reproducibility and validity of a semiquantitative food frequency questionnaire. Am J Epidemiol. 1985;122:51–65
25. Hu FB, Sigal RJ, Rich-Edwards JW, et al. Walking compared with vigorous physical activity and risk of type 2 diabetes in women: a prospective study. JAMA. 1999;282:1433–1439
26. Manson JE, Rimm EB, Stampfer MJ, et al. Physical activity and incidence of non-insulin-dependent diabetes mellitus in women. Lancet. 1991;338:774–778
27. Field AE, Coakley EH, Must A, et al. Impact of overweight on the risk of developing common chronic diseases during a 10-year period. Arch Intern Med. 2001;161:1581–1586
28. . The Expert Committee on the Diagnosis and Classification of Diabetes Mellitus. Report of the expert committee on the diagnosis and classification of diabetes mellitus. Diabetes Care. 1997;20:1183–1197
29. Hu FB, Leitzmann MF, Stampfer MJ, Colditz GA, Willett WC, Rimm EB. Physical activity and television watching in relation to risk for type 2 diabetes mellitus in men. Arch Intern Med. 2001;161:1542–1548
30. Dellavalle RP, Drake A, Graber M, et al. Statins and fibrates for preventing melanoma. Cochrane Database Syst Rev. 2005:CD003697
31. Knowler WC, Fowler SE, Hamman RF, et al.. Diabetes Prevention Program Research Group. 10-year follow-up of diabetes incidence and weight loss in the Diabetes Prevention Program Outcomes Study. Lancet. 2009;374:1677–1686
32. Lindström J, Louheranta A, Mannelin M, et al.. Finnish Diabetes Prevention Study Group. The Finnish Diabetes Prevention Study (DPS): lifestyle intervention and 3-year results on diet and physical activity. Diabetes Care. 2003;26:3230–3236
33. Malik VS, Popkin BM, Bray GA, et al. Sugar-sweetened beverages and risk of metabolic syndrome and type 2 diabetes: a meta-analysis. Diabetes Care. 2010;33:2477–2483
34. Huxley R, Lee CM, Barzi F, et al. Coffee, decaffeinated coffee, and tea consumption in relation to incident type 2 diabetes mellitus: a systematic review with meta-analysis. Arch Intern Med. 2009;169:2053–2063
35. Freeman SR, Drake AL, Heilig LF, et al. Statins, fibrates, and melanoma risk: a systematic review and meta-analysis. J Natl Cancer Inst. 2006;98:1538–1546
36. Robins JM, Hernán MAFitzmaurice G, Davidian M, Verbeke G, Monenberghs G. Estimation of the causal effects of time-varying exposures. In: Longitudinal Data Analysis. Handbooks of Modern Statistical Methods. 2009 Boca Raton, FL Chapman & Hall/CRC:553–599
37. Taubman SL, Robins JM, Mittleman MA, et al. Intervening on risk factors for coronary heart disease: an application of the parametric g-formula. Int J Epidemiol. 2009;38:1599–1611
38. Gooley TA, Leisenring W, Crowley J, et al. Estimation of failure probabilities in the presence of competing risks: new representations of old estimators. Stat Med. 1999;18:695–706
39. Young JG, Cain LE, Robins JM, et al. Comparative effectiveness of dynamic treatment regimes: an application of the parametric g-formula. Stat Biosci. 2011;3:119–143
40. Efron B, Tibshirani R. Bootstrap methods for standard errors, confidence intervals, and other measures of statistical accuracy. Stat Sci. 1986;1:54–75
41. de Koning L, Malik VS, Rimm EB, et al. Sugar-sweetened and artificially sweetened beverage consumption and risk of type 2 diabetes in men. Am J Clin Nutr. 2011;93:1321–1327
42. Aune D, Ursin G, Veierød MB. Meat consumption and the risk of type 2 diabetes: a systematic review and meta-analysis of cohort studies. Diabetologia. 2009;52:2277–2287
43. Robins JM. A new approach to causal inference in mortality studies with a sustained exposure period-application to control of the healthy worker survivor effect. Math Model. 1986;7:1393–1512
44. Willett W, Stampfer MJ, Bain C, et al. Cigarette smoking, relative weight, and menopause. Am J Epidemiol. 1983;117:651–658
45. Fung TT, McCullough M, van Dam RM, et al. A prospective study of overall diet quality and risk of type 2 diabetes in women. Diabetes Care. 2007;30:1753–1757
46. Hernán MA, VanderWeele TJ. Compound treatments and transportability of causal inference. Epidemiology. 2011;22:368–377

Supplemental Digital Content

© 2013 Lippincott Williams & Wilkins, Inc.