Development of Cardiovascular Risk Factors From Ages 8 to 18 in Project HeartBeat!
Study Design and Patterns of Change in Plasma Total Cholesterol Concentration
Background Project HeartBeat! is a longitudinal study of the development of cardiovascular risk factors as growth processes. Patterns of serial change, or trajectories, from ages 8 to 18 years for plasma total cholesterol concentration (TC) and percent body fat illustrate the design and synthetic cohort approach of the study.
Methods and Results Six hundred seventy-eight children (49.1% female, 20.1% black) entered the study at ages 8, 11, and 14 years and were followed up with examinations every 4 months for ≤4 years. Multilevel analysis demonstrated trajectories for population mean values of TC and percent body fat in sex-specific synthetic cohorts from ages 8 to 18 years. Polyphasic patterns of change in TC were confirmed, with notable sex differences in age patterns and with minimum mean values of TC of 3.85 mmol/L for females and 3.59 for males. As illustrated by data for males, the approximate 75th percentile values of mean TC ranged from 4.78 mmol/L at its early peak to 4.06 at its late-teen nadir. Percent body fat exhibited a trajectory closely parallel with that for TC only for males and appeared to be unrelated for females.
Conclusions The polyphasic trajectory for TC from ages 8 to 18 years differs between females and males, indicates marked age variation in 75th percentile values and, in males only, closely parallels the trajectory for percent body fat. These and other results indicate the value of both follow-up every 4 months across age intervals to detect rapid risk factor change and the synthetic cohort approach for gaining new insights into the dynamics and possible determinants of this change from ages 8 to 18 years.
There are well-recognized relations between growth and maturation and cardiovascular disease risk factors in adolescence, stronger than those of these risk factors with chronological age.1 2 3 4 5 6 This suggests that intrinsic aspects of development may be important determinants of risk factor levels and their changes in this period of life. Consequently, there is potential value in the study of development of these risk factors as growth processes. Total cholesterol concentration provides a specific example. Many cross-sectional studies of blood lipids have been conducted in school-age populations. Those studies reporting age-specific mean values by single years of age separately for females and males were previously reviewed and strongly suggested a polyphasic pattern of change in cholesterol concentration during childhood and adolescence.7 8 As we described briefly on the basis of preliminary analysis of the present data,9 total cholesterol concentrations decrease at later ages and decline to a greater degree in males than in females. How this dynamic pattern of change relates to presumed determinants of total cholesterol concentration, the rationale for particular cholesterol screening criteria, or the potential age- or sex-related variation in responsiveness of total cholesterol concentration to interventions are questions that warrant greater attention than they have received.
A clearer understanding of the dynamics of change in risk factors and much closer investigation of their relations to measures of growth and maturation and other covariates could have important theoretical and practical significance. Project HeartBeat! was designed to provide such understanding in relation to each of the major cardiovascular risk factors: blood lipids, blood pressure, and echocardiographic measures and smoking behavior. The present report describes the study design and methods and demonstrates trajectories of change in plasma total cholesterol concentration from ages 8 to 18 years, by sex. In addition, change in cholesterol concentration is examined in relation to change in percent body fat, by sex, across the same age range.
Previous longitudinal studies of children were not designed to evaluate the dynamics of change in cardiovascular risk factors.1 This is chiefly because time intervals of ≥1 year between observations, which are typical of previous studies, provide insufficient frequency of observation. Growth studies in children and adolescents require more frequent examinations, ideally at 3- to 4-month intervals, because of the dynamics of rapid change in the characteristics of interest.
Project HeartBeat! was uniquely designed to study development of the major risk factors in this way. As shown in Fig 1⇓, the design provided observation every 4 months for up to 4 years for each participant. By enrolling cohorts at three age levels, initially 8, 11, and 14 years, and planning for ≤4 years of follow-up in the first phase of data collection, we could obtain observations at common ages during a 1-year age interval for successive cohorts. This overlap in age at observation would occur at age 11 for cohorts 1 and 2 and at age 14 for cohorts 2 and 3. On this basis, a synthetic cohort could be constructed analytically, allowing the assessment of change in each study variable longitudinally from ages 8 to 18 years, with adjustment for possible cohort effects.
The target population consists of residents of The Woodlands and Conroe, Tex. The Woodlands is a planned community in Montgomery County, 30 miles north of Houston, Tex, with a population that at the start of this study was 92% white, 4% Hispanic, 2% black, and 2% other. The city of Conroe is 10 miles north of The Woodlands and is the center of the Conroe Independent School District, in which nearly all study participants in both areas were enrolled. Conroe is 75% white, 13% black, and 12% other (including Hispanics). Although a broad range of socioeconomic levels is present in both communities, middle- to upper-middle-income strata are typical of The Woodlands, and lower- to middle-income strata are more typical of Conroe.
The categories of data collected are listed in Table 1⇓ and include the risk factors themselves (hemodynamics, blood lipids, and smoking behavior) and key covariates (body size and composition, maturation, diet/nutrition, physical activity/fitness, and personal/family/social histories). The frequency of observation for most items was three times per year, ie, every 4 months. For blood pressure, two separate appointments were scheduled within 2 weeks of the initial examination for a total of three occasions at baseline, whereas at subsequent examinations, blood pressure measurements were required on two occasions 1 week apart. Several interview items were assessed only at baseline and anniversary examinations, as was the case for hand-wrist roentgenogram for bone age determination, measurements of urinary electrolytes, and assessment of physical fitness. Cotinine determination was limited to participants ≥14 years old and samples of participants with and without changes in self-reported smoking history. Plasma lipid concentrations were determined in fasting samples obtained and processed in the Lipid Research Laboratory at Baylor College of Medicine. A Cobas Fara II analyzer was used for the enzymatic process of cholesterol determination. For this report, values for plasma total cholesterol concentration, originally determined in units of milligrams per deciliter, were converted to millimoles per liter by multiplication by 0.02586.10 Percent body fat was estimated on the basis of anthropometric and bioelectric impedance data by use of standard methods and the prediction equations of Cameron11 and Guo and colleagues.12
The study protocol was approved by the institutional review committees of The University of Texas–Houston Health Science Center and Baylor College of Medicine. Phase I data collection was conducted from October 1991 through August 1995, with enrollment completed within the first 21 months of data collection. Informed consent or assent and parental consent were documented for each participant. A report of current findings was provided to each family and their designated physician after each examination. The study population was selected to include equal proportions of females and males and ≈20% blacks in numbers sufficient to yield a minimum of 50 remaining participants per age-sex group of nonblacks at the end of 4 years of follow-up. A maximum attrition of 10% per year was assumed for planning purposes. The 8-year-old cohort was overrepresented to ensure sufficient numbers for potential extended individual follow-up of this group through age 18. Table 2⇓ indicates in detail the composition of the study population by sex, cohort, and ethnicity. Overall, 49.1% of participants were female and 20.1% were black. Nonblack minorities were very few in number.
As planned, greater numbers were recruited into cohort 1, particularly for blacks, than into the other two cohorts. An extended effort was required to fulfill the recruitment goal in the project's minority community, particularly for cohorts 1 and 2, as described in a separate report.13 In general, there was a high degree of participation in the scheduled examinations up to the withdrawal date or termination of follow-up for any given participant. The overall dropout rate was much less than assumed in planning. Altogether, during 4 years of data collection, only 153 (22.6%) of 678 participants withdrew from the project after the baseline examination. Group-specific withdrawal rates were 19.9% for blacks and 23.2% for nonblacks, 21.3% for females and 23.8% for males, and 16.2%, 23.9%, and 32.9% for cohorts 1, 2, and 3, respectively. Data in Table 3⇓ indicate that for those who had not officially withdrawn from the project and who therefore remained eligible for examination, completion rates for actual examinations remained very high,≥94%, through the 9th examination. At that point, >64% of the original study population remained in active follow-up and were examined according to schedule. Before many participants became eligible for their 10th examination, the phaseout of data collection began, so that only 80% to 90% of those eligible could be scheduled and actually examined, even though they had not withdrawn from follow-up. The overall mean number of examinations conducted among the 678 members of the study population was 8.3 (data not shown), and of 5809 potential examinations, 5637 (97.0%) were completed. Of these, 5401 examinations (95.8%) included valid determinations of total cholesterol concentration and 5426 (96.3%) included valid determinations of percent body fat.
Analysis of the resulting data was conducted only after very extensive quality assessment with manual and analytic evaluation components, with adjudication by the Project Steering Committee (D.R.L., M.Z.N., R.B.H., and J.A.G.) for each decision concerning raw data. In this process, detailed descriptive statistics were examined for each variable before statistical modeling. For modeling of the patterns of change, or trajectories, of risk factor measures and covariates with age, a general linear mixed-effects model was used. These models were fitted to study data by means of the multilevel statistical modeling approach and MLn software.14 15 16
The MLn regression analysis program computes maximum likelihood estimates of the parameters for mixed linear models. Data from the study are hierarchical, with level 1 of the hierarchy being repeated measurements of the response variables (in the present study, total cholesterol concentration and percent body fat) within subjects and level 2 being the subjects. At level 1, a linear model of the dependent variable y, such as total cholesterol concentration, is expressed as a function of age and other time-dependent covariates separately for each individual study participant using all observations on those variables for that individual. At level 2, the level 1 coefficients are expressed as linear functions of covariates constant within individuals, such as sex. When the level 1 and 2 models are combined, the result is an appropriate general linear mixed model.
This model is similar to an ordinary regression model in that it contains terms expressing the dependence of the outcome variable on a linear combination of predictors, such as sex, age, age2, age3, and age-by-sex interaction. The ability to include quadratic and cubic terms in age in describing the trajectory for total cholesterol concentration was considered necessary due to the hypothesis based on cross-sectional data that this age-related trajectory would be polyphasic in form from ages 8 to 18 years. The model differs from an ordinary regression model in that the “error” term is composed of several parts (hence “mixed”) that reflect the hierarchical nature of the design and the change in variance (and covariance) with selected explanatory variables, such as age of the subject.
A major advantage of the mixed linear model is its ability to account appropriately for the correlations between measurements inherent in a longitudinal study. An additional advantage is the flexibility to model variations in variances and correlations as a function of explanatory variables. Furthermore, the nature of the analysis does not require either equal numbers of observations or identically timed observations for all subjects, as in repeated measures ANOVA. The computational difficulties encountered in fitting mixed linear models to data from longitudinal studies have been overcome to a great extent by the MLn software and other comparable programs.17
The frequency of scheduled examinations (every 4 months), multiple age cohorts, and high level of participation together provided >5400 determinations of total cholesterol concentration at ages ranging from 8 to 18 years. The results for the actual design with thrice-yearly examinations provide a strikingly different picture of cholesterol change than would have been possible had the design been limited to only baseline and anniversary follow-up examinations, as shown in Fig 2a⇓ and 2b⇓. The full series of observations for the first-enrolled participant in cohort 1, selected arbitrarily for illustration, comprises a total of 11 thrice-yearly examinations (upper panel). This series of actual data indicates sharp fluctuations in values, with peaks late in the 8th year and mid- to late-10th year and substantially lower values at other points in the series. Had only the baseline and anniversary data been obtained (lower panel), the pattern for this individual participant would have provided much less insight into the actual variability occurring during this age interval. In the data set as a whole, no evidence was found for seasonal variation (data not shown).
The full data set for all three cohorts was used to model the trajectories for total cholesterol concentration by age and sex. To assess the dependence of the average total cholesterol concentration and its variability on age and sex, a regression model that included terms for age, age2, age3, and sex, as well as age-by-sex interactions, as described, was fitted. Between-subject variances for the constant (intercept) and age terms and the covariance of the age and constant terms were included as well as the within-subject variance. Estimates were calculated by use of MLn statistical software14 as noted above and are shown under “Total Cholesterol” in Table 4.⇓
An ≈5%-level test of the hypothesis that each of the parameters is zero is used to reject the hypothesis whenever the absolute value parameter estimate exceeds 1.96 times the estimated standard error. By this criterion, all fixed coefficients differ from zero except the sex term (because 0.2805/1.932 is <1.96), and thus sex is not likely an important predictor variable for the average level. However, each of the age, age2, and age3 terms is nonzero, which suggests a distinctly nonlinear average trajectory of total cholesterol concentration with age. Furthermore, the presence of nonzero between-subject variances for constant and age indicate that these coefficients vary significantly between subjects. The covariance of the age coefficients and constant terms appears not to differ significantly from zero. The main importance of these terms in this analysis is to model appropriately the complex pattern of variance and covariance induced by the longitudinal nature of the study design. The nonsignificance of the sex term and the significance of each of the sex-by-age interaction terms suggest that the average total cholesterol concentration does not differ significantly between the sexes, but the shapes of the average trajectories for males and females appear to differ, at least to the extent that these are described by a cubic polynomial in age.
To test the dependence of percent body fat and its variability on age and sex, a regression model that included terms for age, age2, age3, sex, and age-by-sex interactions was fitted with the same error structure as the model for total cholesterol concentration. Estimates of the coefficients and their standard errors are displayed in the second pair of columns in Table 4.⇑
In accordance with the same test criterion as before, each of the fixed coefficients, including the interaction terms sex×age, sex×age2 and sex×age3, differs from zero. This result suggests that the trajectories for percent body fat differ in average level as well as shape according to sex. The variance-covariance results are similar to those for total cholesterol concentration, except that covariance of the age coefficients and the constant terms appears to be nonzero.
The predicted values for total cholesterol concentration were calculated by use of the fixed parameter estimates in Table 4⇑. These trajectories are shown in Fig 3⇓. Early in the 8th year of age, mean values were slightly higher for females and increased very slightly but then declined to a nadir at about age 16, followed by increasing values into the late teens. Mean values of total cholesterol concentration for males continued to increase until nearly age 10, then decreased more sharply than for females to reach a nadir nearly 1 year later in age, at almost 17 years. Peak mean values were 4.32 and 4.34 mmol/L at ages 9 and 10 years for females and males, respectively, and minimum values were 3.85 and 3.59 mmol/L at ages 16 and 17 years, respectively. The net decline was 0.47 mmol/L (10.8%) for females and 0.75 mmol/L (17.3%) for males.
To take the more striking pattern for males as the example, the curve of change in absolute values shown in Fig 3⇑ is complemented by the upper bound of its corresponding 50% CI, calculated from the estimated variances, in Fig 4⇓. This confidence bound is presented because it is an estimate of the 75th percentile of the distribution at each age. This percentile from the Lipid Research Clinics data was the basis for the selection of 170 mg/dL (or 4.4 mmol/L) by the National Cholesterol Education Program as the fixed cutpoint in screening to distinguish “borderline” or higher cholesterol values for subjects from ages 2 to 19 years.10 Clearly, actual cholesterol concentrations at the 75th percentile level vary markedly with age for males between 8 and 18 years, from a maximum of 4.78 mmol/L at ≈10 years of age to a minimum of 4.06 at ≈17 years of age. Estimates of the age-specific standard deviation varied from 0.67 to 0.83 mmol/L.
Finally, the same method of calculation of the predicted values for percent body fat from the model in Table 4⇑ reveals the trajectories for percent body fat, by sex, from ages 8 to 18 years, as shown in Fig 5⇓. The trajectory for percent body fat for males peaks at 24.0% at ≈9.5 years of age and decreases sharply past 16 years of age, reaching 14.7% by age 18. The pattern is closely parallel to that for plasma total cholesterol concentration. By contrast, the pattern of change in percent body fat for females indicates only very slight change, between 25.4% and 26.3%, bearing no apparent relation to the trajectory observed for total cholesterol concentration. Age-specific estimates of the standard deviation varied from 7.7% to 10.6%.
We have shown that it is feasible to conduct a study designed like Project HeartBeat!, despite its substantial demand on participant time, with each thrice-yearly visit requiring multiple contacts for full data collection. Recruitment and enrollment goals were met, including those for the minority component from a newly identified study community. To achieve this goal required extension of recruitment time and increased emphasis on the cohort 1 age group of minority participants.13 There was some resulting imbalance in age distributions of blacks and nonblacks, but in fact the Tanner stage ratings18 for sexual maturity at entry were much closer between ethnic groups than if their chronological age distributions had been more closely matched as originally planned (data not shown). This reflects the tendency for sexual maturation to have an earlier onset among contemporary US blacks than among nonblacks (J.M. Tanner, MD, PhD, unpublished data, 1994). Participation and retention were excellent throughout the project, with the attrition rate below that projected for both black and nonblack participants. These initial results from Project HeartBeat! lead to several conclusions.
The full data set for each individual participant is expected to give a much more adequate measure of variation in plasma total cholesterol concentration than if examinations had been conducted only annually, as suggested by the example in Fig 2a⇑ and 2b⇑. Analyses were planned in which trajectories for risk factors and covariates would be estimated within cohorts, at both individual and group levels. The strength of these analyses depended importantly on the additional power resulting from the availability, on average, of >8 observations per participant during the follow-up period of ≤4 years.
For the data set including all participants in the synthetic cohort approach, the applicability of statistical analysis based on the general mixed linear model has been shown. This synthetic cohort approach yielded the anticipated pattern of the average trajectories for total cholesterol concentration across ages from 8 to 18 years for both females and males. The results have confirmed, in direct longitudinal analysis, the polyphasic changes in total cholesterol concentration suggested by previous cross-sectional observations.1 They demonstrate important and previously little-recognized dynamics of change in mean plasma total cholesterol concentration during adolescence and differences in pattern by sex. Future analyses will extend these observations by inclusion of body size and composition, sexual maturation, diet and physical activity, and other covariates.
The initial results presented here already suggest more complex patterns of change in total cholesterol concentration from ages 8 to 18 years than have been generally recognized. These patterns suggest that the fixed screening criterion of 4.40 mmol/L (170 mg/dL) to distinguish “borderline” or higher from “acceptable” levels will identify quite variable proportions of adolescents at different ages as needing repeat measurements or further intervention.10 For males, the fixed 75th percentile cutpoint from the Lipid Research Clinics data for ages 2 to 19 years corresponds approximately to the observed population mean at 10 years of age (half those screened would be considered not acceptable) but corresponds to the 75th or higher percentile at ages ≥13 years (substantially <25% of those screened would be considered not acceptable). This evidence suggests that further evaluation of the effects of fixed and variable screening criteria is warranted.
Furthermore, the contrast between males and females in the apparent relation between trajectories for total cholesterol concentration and percent body fat raises questions as to the role of adiposity as a determinant of blood lipid concentrations in adolescence. The close parallel between trajectories of these variables in males suggests a strong, though not necessarily causal, association between them. Their lack of apparent association in females indicates that adiposity does not have the same role in the variation of total cholesterol concentration in females as in males across this age range and therefore that significant determinants of this major cardiovascular risk factor differ importantly by sex.
These observations point to potential differences by age and sex in responsiveness to interventions intended to modify blood lipids through control of adiposity. Whatever the determinants of change in total cholesterol concentration in adolescence, might these interact differently with interventions at different points in the developmental process for cholesterol (for example, at the peak versus the nadir of cholesterol concentrations by age)? It may be fruitful to reassess the experience of intervention studies to modify total cholesterol concentrations in adolescence by control of adiposity and to test the hypothesis that responses to intervention vary with age and differ in males and females. Support for this hypothesis would have important implications for the design and evaluation of future interventions against elevated total cholesterol concentrations in youth. Such reassessment should also take into account the separate trajectories by sex for each of the major blood lipid components (eg, LDL cholesterol, HDL cholesterol, and triglyceride concentrations).
Finally, it is appropriate to draw an analogy with growth in stature (height). Average trajectories for height at the group level offer only a crude insight into the underlying patterns of change in individuals.18 By contrast, where individual height velocity curves are observed, the phenomenon of the adolescent growth spurt is revealed, and marked interindividual variation in its timing becomes apparent. Thus, detailed characterization of growth curves for individuals reveals variation in “growth tempo” that is only faintly suggested by the average, low-amplitude curves that represent aggregates of individuals. Correspondingly, we expect to describe in future reports interindividual variation in trajectories of each of the major cardiovascular risk factors with their important covariates, taking into account age, sex, ethnicity, sexual maturation, and other factors.
The authors acknowledge with gratitude the contribution of time and dedication of each Project HeartBeat! participant and family. Cooperation of the Conroe Independent School District and generous support of The Woodlands Corporation are deeply appreciated. The Woodlands and Conroe Advisory Committees have assisted greatly in the planning and conduct of this study. We thank Prof James M. Tanner for helpful advice on the design of the study while he was Visiting Professor at the School of Public Health. The authors also wish to acknowledge the essential contributions of the Project HeartBeat! coinvestigators to the design and implementation of this study, including Drs Nancy Ayers, John T. Bricker, John Kirkland, Claudia Kozinetz, Daniel Oshman, Alexander Roche, and William J. Schull. Senior staff of the project for data management and field center management were Tony Arrey and Marilyn Morrissey, and Candace Ayars and Pamela Folsom, respectively. Dr Millicent Higgins served as Scientific Program Administrator for the project under Cooperative Agreement U01-HL-41166, National Heart, Lung, and Blood Institute, which provided major funding for the project. Support of the Centers for Disease Control and Prevention, through the Southwest Center for Prevention Research (U48/CCU609653), and that of Compaq Computer Corporation are also gratefully acknowledged, as is that of The University of Texas–Houston Health Science Center, School of Public Health.
Guest editor for this article was Salim Yusuf, MBBS, DPhil, Hamilton Civic Hospital Research Center, Hamilton, Ontario, Canada.
- Received July 31, 1996.
- Revision received December 4, 1996.
- Accepted January 1, 1997.
- Copyright © 1997 by American Heart Association
Berenson GS. Cardiovascular Risk Factors in Children: The Early Natural History of Atherosclerosis and Essential Hypertension. New York, NY: Oxford University Press; 1980.
Mahoney LT, Lauer RM, Lee J, Clarke WR. Factors affecting tracking of coronary heart disease risk factors in children. In: Williams CL, Wynder EL, eds. Hyperlipidemia in Childhood and the Development of Atherosclerosis. New York, NY: Annals of the New York Academy of Sciences; 1991.
Srinivasan SR. Biologic determinants (or correlates) of serum lipoproteins in children. In: Berenson GS, ed. Causation of Cardiovascular Risk Factors in Children. New York, NY: Raven Press; 1985.
Task Force on Blood Pressure Control in Children. Report of the Second Task Force on Blood Pressure Control in Children: 1987. Pediatrics. 1987;79:1-25.
Labarthe DR, Hanis CL. Obesity, atherosclerosis, and hypertension in youth. In: Filer LJ, Lauer RM, Luepker RV, eds. Prevention of Atherosclerosis and Hypertension Beginning in Youth. Philadelphia, Pa: Lea and Febiger; 1994.
Labarthe DR, O'Brien B, Dunn K. International comparisons of plasma cholesterol and lipoproteins. In: Williams CL, Wynder EL, eds. Hyperlipidemia in Childhood and the Development of Atherosclerosis. New York, NY: Annals of the New York Academy of Sciences; 1991.
Labarthe DR, Harrist RB, Dai S, Grunbaum JA, Nichaman MZ. Study of changes in cardiovascular risk factors in adolescence as growth processes: dynamics of total cholesterol concentration in Project HeartBeat! Am J Epidemiol. 1996;143(suppl):S49. Abstract.
National Cholesterol Education Program: Report of the Expert Panel on Blood Cholesterol Levels in Children and Adolescents. Washington, DC: US Dept of Health and Human Services; 1991. US Dept of Health and Human Services publication NIH 91-2732.
Cameron N. The methods of auxological anthropometry. In: Falkner F, Tanner JM, eds. Human Growth: A Comprehensive Treatise. New York, NY: Plenum Press; 1986.
Guo S, Roche AF, Houtkooper L. Fat-free mass in children and young adults predicted from bioelectric impedance and anthropometric variables. Am J Clin Nutr. 1989;50:435-443.
Grunbaum JA, Labarthe DR, Ayars C, Harrist R, Nichaman MZ. Recruitment and enrollment for Project HeartBeat!: achieving the goals of minority inclusion. Ethn Dis. In press.
Goldstein H. Multilevel Statistical Models. 2nd ed. London, UK: Edward Arnold; 1995.
Harrist RB, Kimball KT. Multilevel analysis for a mixed longitudinal design. In: Hauspie R, Lindgren G, Falkner F, eds. Essays in Auxology. Welwyn Garden City, UK: Castlemead Publications; 1995.
Rabash J, Woodhouse G. MLn Command Reference, Version 1.0. London, UK: Multilevel Models Project, Institute of Education, University of London; 1995.
Tanner JM. Foetus into Man: Physical Growth from Conception to Maturity. 2nd ed. Ware, UK: Castlemead Publications; 1989.