RSS Feeds
Two Risk-Scoring Systems for Predicting Incident Diabetes Mellitus in U.S. Adults Age 45 to 64 Years
- Henry S. Kahn, MD;
- Yiling J. Cheng, MD, PhD;
- Theodore J. Thompson, MS;
- Giuseppina Imperatore, MD, PhD; and
- Edward W. Gregg, PhD
Abstract
Background: Simple prediction scores could help identify adults at high risk for diabetes.
Objective: To derive and validate scoring systems by using longitudinal data from a study that repeatedly tested for incident diabetes.
Design: Prospective cohort, divided into derivation and validation samples.
Setting: The ARIC (Atherosclerosis Risk in Communities) study, which followed participants for 14.9 years beginning in 1987 to 1989.
Participants: 12 729 U.S. adults (baseline age, 45 to 64 years; 22.8% black). Follow-up was 96.1% at 5 years and 72.2% at 10 years.
Measurements: Anthropometry, blood pressure, and pulse (basic system) plus a fasting blood specimen assayed for common analytes (enhanced system). Diabetes was identified in 18.9% of participants. Risk score integer points were derived from proportional hazard coefficients associated with baseline categorical variables and quintiles of continuous variables.
Results: The basic scoring system included waist circumference (10 to 35 points); maternal diabetes (13 points); hypertension (11 points); and paternal diabetes, short stature, black race, age 55 years or older, increased weight, rapid pulse, and smoking history (≤8 points each). The enhanced system included glucose (6 to 28 points); waist circumference (5 to 21 points); maternal diabetes (8 points); and triglycerides, black race, paternal diabetes, low high-density lipoprotein cholesterol concentration, short stature, high uric acid, age 55 years or older, hypertension, rapid pulse, and nonuse of alcohol (≤7 points each). When applied to the validation sample, ascending quintiles of the basic system were associated with a 10-year incidence of diabetes of 5.3%, 8.7%, 15.5%, 24.5%, and 33.0%, respectively. Quintiles of the enhanced system were associated with a 10-year incidence of 3.5%, 6.4%, 11.5%, 19.3%, and 46.1%.
Limitations: The risk scoring systems had no question regarding previous gestational diabetes, and knowledge of parental diabetes may be uncertain. The analyzed cohort was restricted by age and race; the systems may be less effective in other samples.
Conclusion: Basic information identified adults at high risk for diabetes. Additional data from fasting blood tests better identified those at extreme risk.
Primary Funding Source: Centers for Disease Control and Prevention.
Editors' Notes
Context
-
Prediction scores that combine multiple factors could help identify people who are at risk for diabetes.
Contribution
-
Using data from a large prospective study, these investigators derived and tested scoring systems for predicting development of diabetes within 10 years among middle-age adults. A basic system, which included such items as waist circumference, parental diabetes, hypertension, short stature, and black race, identified groups with varying risk for diabetes. An enhanced system that included such blood tests as fasting glucose, lipids, and uric acid better identified those at highest risk.
Caution
-
These scoring systems need to be validated in different populations and their utility evaluated before widespread application.
—The Editors
The incidence of diabetes among U.S. adults continues to increase (1, 2). Against this disturbing trend, good evidence (3, 4) now indicates that dietary and exercise interventions offered to nondiabetic adults at high risk can materially reduce or delay their conversion to type 2 diabetes, and a recent study (5) indicates that such interventions are most effective among people with the highest a priori risk for diabetes. These interventions, however, may be burdensome to some patients or their families, and the substantial societal costs of these interventions could eventually be borne by insurers or public health agencies (6, 7). For these reasons, it may be difficult to justify providing expensive intervention programs to people who are at relatively low risk for diabetes. Clinicians who provide medical care to adults need simple, validated tools with which to identify those most likely to benefit from participation in a diabetes prevention program.
Research protocols for estimating a person's future risk for type 2 diabetes have depended primarily on identifying impaired glucose tolerance through a 2-hour oral glucose tolerance test (8–11). This 2-hour challenge test, however, is relatively costly, inconvenient, and labor-intensive, and its reliability has been questioned (12, 13).
We sought to develop simple scoring systems that could efficiently identify adults at high risk for diabetes without requiring the 2-hour challenge test. To do so, we used 14.9 years of longitudinal data from the ARIC (Atherosclerosis Risk in Communities) Study to derive and validate a basic diabetes prediction scoring system that depended on each participant's baseline anthropometric characteristics, sex, parental history of diabetes, and other clinical variables but did not require a blood specimen. We also derived and validated an enhanced scoring system that evaluated the same variables as the basic system along with baseline concentrations of glucose and other analytes that are commonly assessed in a fasting blood sample.
Methods
Population Sample
The ARIC Study used probability sampling of white and black adults in 4 U.S. communities to assemble a population-based cohort of 15 792 adults age 45 to 64 years at baseline in 1987 to 1989 (14–16). To comply with the requirements of human subject research review committees, ARIC participants provided written consent for their involvement and data sharing. We obtained public-use, limited-access data released by the Collaborative Studies Coordinating Center of the University of North Carolina at Chapel Hill (17) for our analysis. We excluded 60 persons who did not consent to sharing their information. We excluded another 3003 persons because they had prevalent diabetes at baseline or because we lacked complete data (Figure 1).
ARIC = Atherosclerosis Risk in Communities.
Assessment of Risk Factors and Outcomes
We chose candidate baseline variables for their simplicity, low cost, common availability, and mention in previous models of diabetes risk. At the baseline interview, the diabetes status of participants' natural parents was determined from their response to a question about whether either parent had ever had “diabetes (sugar in the blood).” Racial group was categorized as either nonblack (reference, of whom <0.4% were other than white) or black (16); the public use data set provided no information on ethnic descriptions. Baseline age was categorized initially in 5-year intervals, beginning with 45 to 49 years (reference). Smoking history was categorized as none (reference), moderate (<24 cigarette pack-years or a similar exposure), or heavy (≥24 cigarette pack-years or similar exposure) on the basis of responses to extensive questions about lifetime exposure to cigarettes, pipes, and cigars or cigarillos. Alcohol consumption status was categorized as current drinker (reference), former drinker, or never drinker. Educational attainment was categorized as high (17 to 21 years; reference), moderate (12 to 16 years), or low (<12 years).
Height and weight were measured at the baseline examination without shoes or heavy clothing. Waist circumference (standing) was measured once at the level of the umbilicus while the examiner held the tape horizontally and instructed the participant to breathe quietly. Systolic and diastolic blood pressure values were the average of the second and third readings obtained in a seated position after a 5-minute rest; participants were considered to have hypertension if they had a systolic value of at least 140 mm Hg or a diastolic value of at least 90 mm Hg or reported recent use of antihypertensive medication. Pulse rate was determined from a supine electrocardiogram obtained at rest. Central laboratories performed analyses on baseline fasting serum and plasma specimens by using conventional assays (17). The ARIC Lipid Laboratory participated in the Centers for Disease Control and Prevention–National Heart, Lung, and Blood Institute Lipid Standardization Program. No glycated hemoglobin assay or glucose tolerance test was performed at the baseline examination.
Diabetes status was reassessed primarily at triennial follow-up encounters that included fasting blood specimens (year 3 [in 1990 to 1992], year 6 [1993 to 1995], and year 9 [1996 to 1998]). The year 9 follow-up examination also provided a 2-hour oral glucose (75 g) tolerance test. Participants were considered to have diabetes if they reported a history of physician-diagnosed “diabetes (sugar in the blood)” or if their fasting glucose level was at least 7.0 mmol/L (≥126 mg/dL), their nonfasting glucose level was at least 11.1 mmol/L (≥200 mg/dL), or their 2-hour glucose value at the year 9 follow-up was at least 11.1 mmol/L (≥200 mg/dL). Additional cases of incident diabetes were identified by criterion-based abstractions of hospital records (obtained through surveillance of hospital admissions) and from responses to annual postal or telephone questionnaires distributed during follow-up years 11, 12, 13, and 14 (17).
Statistical Analysis
From the eligible study population of 12 729 participants, we randomly selected a 75% derivation sample (9587 persons) in which to model the incidence of diabetes in relation to baseline risk factors. We used the LIFEREG procedure in SAS, version 9.1 (SAS Institute, Cary, North Carolina) to create Weibull proportional hazard regression models for interval-censored outcome data because the ARIC records provided only an interval for each diabetes diagnosis rather than a precise date. We categorized all continuous variables into quintiles (sex-specific for all except fasting glucose) so that the estimated contribution of these variables to diabetes risk could be expressed through simplified point scores assigned to each of the quintile categories. For anthropometric variables, we used weight, height, and waist circumference rather than complex indices, such as body mass index or waist-to-height ratio.
We then identified a best basic predictive model (without baseline blood information) and a best enhanced predictive model (which included data from analysis of a baseline fasting blood specimen). For each variable in these models, we considered adjacent categories of the variable to be overlapping if their estimated Weibull β coefficients were within 1 SE of each other. We joined any overlapping categories into a single combined categorical term, which we retained if it contributed to the model (P < 0.05). For the 2 final models, each retained categorical variable was represented by a point score proportional to the estimate of its Weibull β coefficient. We standardized the point score values so that the maximum total score equaled 100 and rounded each point score assignment to the nearest integer.
We then evaluated the performance of these best basic and enhanced scoring systems in a validation sample that comprised the remaining 25% of the eligible ARIC cohort (3142 participants). For these evaluations, we used interval-censored Weibull models to estimate the 10-year diabetes incidence rates. Applying each scoring system, we calculated the area under the receiver-operating characteristic (ROC) curve on the basis of the 10-year incidence outcome (18). Through repetitive sampling and replacement of participants in the validation sample, we generated a bootstrapped estimate of the confidence limits for our areas under the ROC curves (19). We also identified the empirical quintiles of basic or enhanced score values and estimated the 10-year diabetes incidence among members of the validation sample with scores in each of the 5 quintiles by using the LIFETEST procedure of SAS with the Kaplan–Meier survival function.
For the purpose of comparison with our derived scoring systems, we used the same validation sample to test 2 alternative simple scoring systems (20, 21) that similarly used integer points for predicting diabetes.
Role of the Funding Source
The National Heart, Lung, and Blood Institute funded the ARIC Study, including its participant follow-up. The Centers for Disease Control and Prevention supported the secondary analysis of ARIC data. Neither funding source had a role in the design of our secondary analysis, its interpretation, or the decision to submit the manuscript for publication.
Results
Cohort at Baseline and Follow-up
Our baseline sample comprised 12 729 participants (mean age, 53.9 years). Of these, 44.5% were male, 22.8% were black, 30.7% had hypertension, 15.1% described a maternal history and 8.8% described a paternal history of diabetes, 29.7% reported moderate and 28.5% reported heavy historical tobacco use, 23.8% reported never using and 17.3% reported formerly using alcohol, and 20.9% reported having fewer than 12 years and 41.7% reported having 12 to 16 years of education. Table 1 shows the quintile cut-points for the distribution of continuous anthropometric and laboratory values.
Statistical follow-up of ARIC participants continued until they were last assessed for incident diabetes. For this purpose, follow-up was 96.1% complete at 5 years and 72.2% complete at 10 years; 2407 (18.9%) of the baseline cohort members were found to have diabetes at some point during follow-up (Figure 1).
Derivation of Prediction Scores
In our derivation sample of 9587 adults, we found a statistically significant age effect only between age 45 to 54 years (revised reference category) and 55 to 64 years. Although men had a slightly higher cumulative diabetes incidence rate than women over 14.9 years (19.4% vs. 18.6%), this sex difference was statistically nonsignificant (P > 0.2) in proportional hazard models that adjusted for baseline age, race, parental diabetes, and hypertension. Former and never drinking were each associated with a slightly elevated diabetes risk, as was lower educational attainment, but these variables did not contribute to our basic model. Variables that described the contributions of height, pulse, weight, and smoking to diabetes risk had overlapping effects of some adjacent risk categories, and these overlaps required us to combine the categories. Figure 2 shows the terms we retained in our basic model (P < 0.036 for all), along with their associated integer points. On the basis of these assigned point values, the derivation sample had a mean score of 38.1 (SD, 18.1) and a median score of 38; the quintile cut-points for the basic scoring system were 21, 33, 43, and 55.
The model does not require blood analysis data. Developed from a derivation sample of 9587 persons age 45 to 64 years who did not have diabetes at baseline.
When we introduced quintile categories for commonly available fasting blood measurements, we found no effects associated with concentrations of total cholesterol, low-density lipoprotein cholesterol, or magnesium. When we included terms for glucose, triglycerides, high-density lipoprotein cholesterol, and uric acid, the terms for sex, tobacco use, educational level, and weight became statistically nonsignificant. In the enhanced model, alcohol use demonstrated overlapping effects associated with no or former alcohol use. We also found overlapping adjacent categories for height; pulse; and concentrations of glucose, triglyceride, high-density lipoprotein cholesterol, and uric acid. Figure 3 shows the terms we retained in our enhanced model after we combined the overlapping categories (P < 0.020 for all). On the basis of their assigned point values, the derivation sample had a mean score of 33.7 (SD, 18.2); a median score of 32; and quintile cut-points of 17, 27, 37, and 50.
The model includes data from a fasting blood sample. Developed from a derivation sample of 9587 persons age 45 to 64 years who did not have diabetes at baseline. HDL = high-density lipoprotein.
Plots of the −log(survival) curves for quintiles of the basic and enhanced score values confirmed that both scoring systems met the proportional hazard assumptions.
Validation of Prediction Scores
When applied to the validation sample of 3142 participants, our basic scoring system had an area under the ROC curve (Figure 4) of 0.71 (95% CI, 0.69 to 0.73). The maximum value of (sensitivity plus specificity) was achieved at a basic score of 38 (sensitivity 69%, specificity 64%). We estimate that the 10-year diabetes incidence was 17.7% among members of our independent validation sample. Ten-year diabetes incidence estimates associated with ascending quintiles of the basic score were 5.3%, 8.7%, 15.5%, 24.5%, and 33.0% (Figure 5).
The ARIC-derived basic system is compared with the DESIR clinical diabetes risk score (20) (neither required a blood sample), and the ARIC-derived enhanced system is compared with the Framingham simple clinical model (21) (both required a fasting blood sample). ARIC = Atherosclerosis Risk in Communities; DESIR = Epidemiologic Study on the Insulin Resistance Syndrome.
Our enhanced scoring system in the validation exercise had an area under the ROC curve of 0.79 (CI, 0.77 to 0.81) (Figure 4). The maximum value of (sensitivity plus specificity) was achieved at an enhanced score of 38 (sensitivity, 74%; specificity, 71%). Ten-year incidence estimates by enhanced risk score quintiles were 3.5%, 6.4%, 11.5%, 19.3%, and 46.1% (Figure 5).
Table 2 summarizes 6 additional scoring systems that have assigned integer points for predicting incident diabetes. Four of these alternative simple systems require information or an age range not available from the ARIC baseline assessments. Of the remaining 2 simple systems, one (20) could be compared with our basic system (no blood specimen required) and the other (21) could be compared with our enhanced system (blood specimen required). When we applied them to the same validation sample, we found that our basic and enhanced systems had better ROC curves when used to identify middle-age U.S. adults at risk for incident diabetes (Figure 4 and Table 2).
Discussion
Through secondary analysis of observational data from 75% of the ARIC cohort, we derived 2 scoring systems that predict incident diabetes among middle-age U.S. adults. We tested both scoring systems in an independent sample from the ARIC cohort and confirmed their utility for identifying persons at high risk for diabetes during a decade of follow-up. Perhaps the most appealing feature of our 2 scoring systems is their ease of use in a clinical or public health setting. Through the adoption of integer point scores, their algorithms are simple enough to be calculated using only a pencil and paper.
Predictive models for diabetes are not new, but few of them have been validated through prospective testing in an independent population. Our analysis builds on an earlier validated study (22) that predicted incident diabetes among members of the ARIC cohort; however, that risk-prediction system was based on only 9 years of follow-up data from fewer than 8000 participants, who were equally divided at baseline into derivation and validation samples. Other than this earlier analysis of ARIC data, we know of only 2 prospective diabetes risk-prediction systems based on independent samples of U.S. adults for which validations have been reported: a validation report (23) that was limited to a small sample of Japanese Americans and a 5-year follow-up study of U.S. adults who were at least 70 years of age at baseline (24). Our literature review through November 2008 found additional reports from Europe, Mexico, and Thailand (20, 25–31) that described the independent validation of diabetes prediction models among community-based samples of middle-age adults during follow-up periods of 3 to 9 years. Of these validated scoring systems, only 4 (20, 24, 25, 27) were constructed with algorithms that were simplified through the use of integer point values.
Beyond their potential for clinical or public health application, validated prospective prediction models may help illuminate the cause of diabetes. For example, in both our basic and enhanced scoring systems, the waist circumference variable contributed more points to risk discrimination than did the weight and height variables combined. This finding is consistent with previous reports based on ARIC data (22, 32), as well as with those from other prospective studies (33–39). Also consistent with previous results (36, 40), we found that short stature was independently associated with diabetes risk in our enhanced model but weight was not. A recent large cohort study (41) found that underweight and overweight status were both positively associated with diabetes risk among older adults. These findings raise the possibility that the often-described association between body mass index (weight divided by height squared) and incident diabetes might be driven as much by reduced height (in the denominator) as by increased weight (in the numerator).
Having a parent with diabetes contributed strongly to diabetes risk in our 2 models as well as in previous prospective studies (20–22, 27, 42, 43). These familial associations possibly involve genetic, epigenetic, environmental, and behavioral pathways. Knowledge of a parent's diabetes status, however, is necessarily incomplete because of inadequate diagnostic opportunities, the parent's premature death, or loss of contact. We found that more ARIC participants were aware of having a mother with diabetes (15.1%) than of having a father with diabetes (8.8%) and that the independent contribution of maternal diabetes was greater than that of paternal diabetes. The relatively greater awareness of maternal diabetes may in part be related to increased screening during pregnancy and the rising prevalence of gestational diabetes among pregnant women (44). A woman whose pregnancy was complicated by gestational diabetes is herself at greatly increased risk for type 2 diabetes (43, 45, 46), but the ARIC data set did not provide any baseline information specifically about gestational screening or diabetes diagnosed in pregnancy. Despite the inherent ambiguities in ascertaining a diabetes history for either parent or for women who have been pregnant, these variables can provide valuable information for any predictive model. Future systems for predicting diabetes should probably include specific questions that relate to diabetic screening and diagnosis during pregnancy.
The inclusion of hypertension in diabetes prediction models is not new (21, 22, 47–49). European studies (50, 51) have shown an elevated resting heart rate to be another risk factor for incident diabetes, but our study may be the first to demonstrate that hypertension and rapid heart rate can contribute simultaneously in the same diabetes prediction model. It is unclear whether these hemodynamic characteristics actively influence diabetes risk or are merely reflections of shared underlying humoral and physiologic factors. Nevertheless, their role as risk markers should now be well established.
Black race contributed independently to our predictive models. This is consistent with previous diabetes incidence studies from the United States (16, 22, 52) and a recent study that showed a higher prevalence of diabetes among black persons than among white persons in the United States (53). After adjustment for possible racial differences in anthropometry, parental history, and some physiologic covariates, the persistence of higher diabetes incidence for black persons suggests that other variables may explain the apparent racial risk factor. These unmeasured variables may include a differential distribution of genetic polymorphisms in black and white populations, greater exposure to adverse environments in early life, or the reduced availability of vitamin D in dark-skinned populations (54, 55).
Smoking and nonuse of alcohol contributed modestly to diabetes risk in our models. Although the effects of these variables showed consistent trends in the expected directions, the strength of these factors did not always reach a level sufficient to justify inclusion in our simple models. We did not confirm the dose–response smoking effect reported by a meta-analysis of 25 prospective studies (56). Our enhanced model, but not our basic model, included an increased risk for diabetes associated with being a nonuser of alcohol at the time of the baseline assessment, a finding that confirms several previous epidemiologic reports (57–60). This consistent effect may reflect the ability of moderate alcohol consumption to improve insulin sensitivity, adiponectin levels, and lipid profiles (61). Future investigators may clarify why smoking was statistically significant only in our basic model, whereas nonuse of alcohol was statistically significant only in our enhanced model.
In our enhanced scoring system, the fasting glucose concentration at baseline was the single largest contributor (up to 28 points; Figure 3) to diabetes risk. Other reports (21, 22, 24) have similarly shown that fasting glucose is a strong predictor of incident diabetes. This is not unexpected, given that people whose fasting glucose concentration is already close to the diagnostic threshold for diabetes are relatively more likely to cross the threshold in the near future. Although high glucose concentrations are associated inherently with increased diabetes risk, we found a similar degree of increased risk associated with the combination of waist circumference (up to 21 points) and fasting triglyceride concentration (up to 7 points). This finding supports the emerging view that type 2 diabetes may reflect not so much an isolated impairment of glucose regulation but rather the complex metabolic consequences of accumulating ectopic lipids (62, 63) or hepatic fat (64). The contribution of increased triglyceride concentration to prospective diabetes risk has been described (21, 22, 24, 51, 59, 65–67), as have the contributions of high uric acid concentration (42, 68–70) and low high-density lipoprotein cholesterol concentration (21, 22, 42, 59, 71).
Our study has limitations. Our scoring systems were derived and validated in a recent cohort of U.S. residents and thus might be suitable to use in similar clinical or epidemiologic settings. However, we have not confirmed their accuracy in predicting diabetes risk for people who are younger than 45 years, older than 64 years, or from racial or ethnic groups other than black or white. This represents a major limitation to the use of our findings. Similarly, the scoring systems derived from cohorts with distinct age or racial or ethnic composition (such as the DESIR [Epidemiological Study on the Insulin Resistance Syndrome] or Framingham studies) may have had reduced utility in predicting incident diabetes in the ARIC validation sample. However, validation of our simple scoring systems could be attempted at relatively low cost in various other samples. In the interest of promoting more validation attempts, we have included 2 more refined scoring systems that we derived from the full sample of eligible ARIC participants (Appendix and Appendix Figures 1 and 2) rather than from only the 75% derivation sample.
See the Appendix for further information. The model does not require blood analysis data. Developed from the full derivation sample of the Atherosclerosis Risk in Communities Study: 12 729 persons age 45 to 64 years who did not have diabetes at baseline.
See the Appendix for further information. The model includes data from a fasting blood sample. Developed from the full derivation sample of the Atherosclerosis Risk in Communities Study: 12 729 persons age 45 to 64 years who did not have diabetes at baseline. HDL = high-density lipoprotein.
Although either of our scoring systems might be tested and used in various settings related to diabetes prevention, we envision an evolving clinical acceptance of our basic scoring system as a low-cost, clinical approach to the early identification of adults who might benefit from interventional programs. Our basic system, which requires no blood samples, could be used to identify asymptomatic persons who should be invited to a more thorough examination that includes blood testing. The basic system would probably also identify persons who already have diabetes and those who are at elevated risk for associated cardiovascular diseases. Sequential application of the basic and enhanced scoring systems, possibly in conjunction with further cardiovascular assessments (72, 73), would help to bring adults at greatest need into appropriate modes of preventive care. Outside the clinical setting, the basic scoring system for diabetes risk could also serve an epidemiologic purpose. We anticipate that insurers or public health agencies could develop survey instruments that use such a system to optimize their allocation of preventive medicine resources among the communities they serve.
Appendix: Alternative Scoring Systems Derived From a Larger Population
We present here 2 alternative diabetes scoring systems, basic and enhanced (Appendix Figures 1 and 2, respectively), that we derived from the largest available sample of nondiabetic participants (12 729 participants at baseline) followed within the ARIC (Atherosclerosis Risk in Communities) Study. Whereas we created the scoring systems presented in our article from a smaller, randomly chosen derivation sample (9587 participants at baseline), the scoring systems presented here take advantage of a larger sample to generate models that are slightly more refined. We used the same procedures to create these more refined models as for the initial derivation sample. Unlike the scoring systems in our article, however, we did not validate these more refined models in an independent sample.
On the basis of the point values assigned in the refined basic model, the mean diabetes risk score for nondiabetic ARIC participants was 37.6 (SD, 17.4); the median score was 37; and the quintile cut-points were 21, 32, 43, and 54. For the refined enhanced model, the mean diabetes risk score for nondiabetic ARIC participants was 32.8 (SD, 18.4); the median score was 30; and the quintile cut-points were 16, 25, 36, and 49.
Article and Author Information
-
Disclaimer: The findings and conclusions in this report are those of the authors and do not necessarily reflect the opinions or views of the ARIC Study or the National Heart, Lung, and Blood Institute, nor do they necessarily represent the official position of the Centers for Disease Control and Prevention.
-
Acknowledgment: The authors thank the ARIC investigators, coordinating center, and volunteer participants and the staff of the National Heart, Lung, and Blood Institute, who authorized sharing of the limited-access longitudinal data sets.
-
Financial Support: By the Centers for Disease Control and Prevention.
-
Potential Financial Conflicts of Interest: None disclosed.
-
Reproducible Research Statement: Study protocol (ARIC): Available at http://www.cscc.unc.edu/aric/index.php. Statistical code (secondary analysis): Available from Dr. Kahn (hkahn{at}cdc.gov). Data set (ARIC): Available through a limited-access distribution agreement (http://www.cscc.unc.edu/aric/utility/docfilter.php?study=aric&filter_type=datadist).
-
Requests for Single Reprints: Henry S. Kahn, MD, CDC Mail Stop K-10, 4770 Buford Highway Northeast, Atlanta, GA 30341; e-mail, hkahn{at}cdc.gov.
-
Current Author Addresses: Drs. Kahn, Cheng, Imperatore, and Gregg and Mr. Thompson: Centers for Disease Control and Prevention, CDC Mail Stop K-10, 4770 Buford Highway Northeast, Atlanta, GA 30341.
-
Author Contributions: Conception and design: H.S. Kahn, Y.J. Cheng, G. Imperatore.
-
Analysis and interpretation of the data: H.S. Kahn, Y.J. Cheng, G. Imperatore, E.W. Gregg.
-
Drafting of the article: H.S. Kahn.
-
Critical revision of the article for important intellectual content: H.S. Kahn, Y.J. Cheng, G. Imperatore, E.W. Gregg.
-
Final approval of the article: H.S. Kahn, Y.J. Cheng, G. Imperatore, E.W. Gregg.
-
Provision of study materials or patients: H.S. Kahn, Y.J. Cheng.
-
Statistical expertise: Y.J. Cheng, T.J. Thompson.
-
Collection and assembly of data: H.S. Kahn, Y.J. Cheng.
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.↵
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.↵
- 27.↵
- 28.↵
- 29.↵
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.↵
- 36.↵
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.↵
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.↵
Related Clinical Content
