Identification of Children’s BMI Trajectories and Prediction from Weight Gain in Infancy

Abstract

Objective:

The goal of this study was to identify patterns of BMI changes across childhood (ages 24 months to 13 years) and to assess whether demographic characteristics, birth weight, and percent infant weight gain from birth to 15 months predicted BMI patterns.

Methods:

Eleven waves of data from the Study of Early Child Care and Youth Development were used. Trained technicians assessed children’s weight at birth and 10 times from 15 months to eighth grade (N = 1364). Latent growth modeling was used to estimate BMI change trajectories, and logistic regression was used to predict membership in trajectory classes.

Results:

Children in the high-rising and low-to-high BMI patterns had the highest BMI of all trajectory groups during middle childhood. Birth weight and infant weight gain were stronger predictors of trajectory membership than gender or race/ethnicity. Infant weight gain predicted high-rising membership over and above the effect of birth weight. African American children had lower birth weight, faster infant weight increase, and higher odds of being in one of the rising trajectories. Risk algorithms are provided.

Conclusions:

Clinicians should monitor weight gain during infancy independent of birth weight. Researchers should continue investigating the lasting physiological effects of early rapid weight gain in infancy.

Introduction

The etiology of childhood obesity has remained unclear in part because patterns of normative weight change in childhood are unclear. Given the risks that childhood obesity and early rapid weight gain appear to exert on adult outcomes such as cardiovascular health and premature death (1,2), among others, understanding common patterns of change and the predictors of such change could considerably advance early detection and intervention. Several efforts employing latent class analysis have identified distinct patterns of children’s weight or BMI change over time (3–6). We took such models a step further by investigating how common assumptions related to variability influence the number and shape of trajectories identified (7). Our goal was to differentiate coherent trajectories that might include patterns with few individuals but important health implications, such as children who might have low BMI in early childhood and thus not seem at risk, but who experience steep or steady increases in BMI in childhood and later develop obesity.

We also examined the relative roles of birth weight and rapid infant weight gain (from birth to 15 months) in the development of childhood BMI and obesity. High birth weight has predicted obesity in early childhood (8,9), but this finding has not been replicated in large studies with longer time horizons (10). Rapid weight gain in the first 6 months to 1 year of life, particularly following low birth weight, has generally been found to predict obesity in early childhood (11–13); the challenge is then to address infants with low birth weight, for whom rapid early weight gain may constitute a catch-up with their same-aged peers (14,15). Inconsistent findings may be a result of modeling entire sample populations together, omitting patterns of change, or assessing predictors one at a time, which ignores potential confounds among them. We examined whether data easily accessible to pediatricians in the first 15 months of life—birth weight for gestational age, early weight increase, race/ethnicity coding, and gender—predicted children’s risk for following trajectories leading to elevated BMI.

Methods

Participants

A secondary data analysis using all four phases of the National Institute of Child Health and Human Development’s Study of Early Child Care and Youth Development (SECCYD) data set supported this investigation. In 1991, a team of researchers enrolled 1,364 children and families from 10 different sites across the nation. Their sampling strategy focused on ensuring a sufficient size sample for factors of primary interest, for example, single-parent families and minority representation. Analyses have shown that the data collected reflect the natural distributions of these characteristics in the catchment and in the 1990 census data, but as a true probability sampling design was not used, the SECCYD data cannot be interpreted strictly as representative of the general population (16). Detailed information on study design, procedures, population characteristics with respect to weight variables, and original institutional review board oversight is available elsewhere (16,17). We requested and received access to this restricted data set from the Inter-university Consortium for Political and Social Research at the University of Michigan.

Outcomes

BMI percentiles at nine waves were the outcomes for this study. Medical professionals or trained technicians assessed children’s height (or length) and weight at each study wave from birth to ninth grade. BMI percentile as assessed against the Centers for Disease Control and Prevention (CDC) 2000 growth charts was the modeled value, as raw BMI is not standardized for growth in early life and rises monotonically (18). We did not model BMI percentile before 2 years of age, as the CDC recommends against using it in a clinical context (18). The final time point, ninth grade, was excluded in this investigation, as it is used in ongoing research for final BMI and other health outcomes we wish to predict using trajectory membership. Therefore, BMI percentiles at nine time points (24, 36, and 54 months and first, third, fifth, sixth, seventh, and eighth grades) were used for trajectory modeling (typical ages for these grades are 6, 8, 10, 11, 12, and 13 years). Overweight and obesity were defined as ≥85th percentile and ≥95th percentile BMI, respectively (18).

Predictors

Birth weight for gestational age was calculated as raw birth weight adjusted by the modeled relation between birth weight and gestational age for the entire data set. We characterized percent weight gain from 0 to 15 months as the ratio of the difference in weight from 0 to 15 months to birth weight for gestational age × 100. The original National Institute of Child Health and Human Development study had retrieved gender as well as race from birth certificates; the race categories were American Indian/Eskimo/Aleutian, Asian/Pacific Islander, Black or African American, White, or Other. A binomial indicator designated Hispanic ethnicity. Although the meaning and utility of such racial and ethnic variables in explaining biological phenomena remain unclear (19), we nevertheless used them as they were elected by parents from the list of available codes provided in the study.

Statistical analysis

We first assessed patterns of missingness, outlying data points, and sample characteristics. To explore patterns of BMI change over time, we used latent growth modeling, a family of analyses that estimates latent variables, or “growth factors”—intercept, slope, quadratic, and cubic terms—that characterize individuals’ change over time. Two methods of classifying unconditional (strictly empirically derived, without predictors) latent growth curves were tested: growth mixture modeling (GMM) and a more restrictive subtype known as latent class growth analysis (LCGA). Both methods accommodated missing data points by full information maximum likelihood, which estimates a likelihood function for each individual based on the available measurements.

GMM identifies distinct groups of individuals who share similar starting values and growth characteristics, such that each group has its own growth model (20,21). An important feature of GMM is that it allows individuals to vary around the mean growth trajectory for their class. This method is computationally intensive but allows greater heterogeneity within the class and typically results in fewer classes being identified. In contrast to GMM, LCGA fixes the variance and covariance estimates for the growth factors and thus assigns each individual in the class the same growth factors (22). This method of identifying patterns of growth curves is simpler, but because individual trajectories are required to be more similar in order to be classified together (essentially having zero variance from the average growth parameters for the class), LCGA typically yields more classes than GMM. LCGA may be used as a step prior to conducting GMM, as distinct classes can be identified and assessed for whether their differences are meaningful enough to warrant keeping them separated. If not, some may be allowed to merge, using the GMM parameterization (6).

The GMM and LCGA models were conducted in Mplus v7.2 (Muthén & Muthén, Los Angeles, California). Growth trajectory class solutions were compared using standard measures of relative fit (see Latent Growth Modeling in online Supporting Information). Evaluation of trajectory characteristics and multivariate multinomial logistic regressions predicting the risks of trajectory membership were conducted in the statistical computing environment R (R Foundation for Statistical Computing, Vienna, Austria). Testing predictors together in a single regression model helped alleviate confounds such as gender with birth weight. Final membership prediction models were determined by the statistical significance of individual predictors at a type I error rate α = 0.05 and model χ² difference tests.

Results

Characteristics of the sample

The data set was analyzed for attrition, patterns of missingness, and outliers that might have resulted in some time points or demographic characteristics being overrepresented or otherwise skewing results. No significant impacts were discovered (see Attrition, Missingness, and Outliers in online Supporting Information for results and accommodations made). Table 1 summarizes the tested weight predictors by race and gender. Only participants born between 37 and 41 weeks’ gestation and who had at least two time points of data were included. Four individuals were excluded following evaluation of the growth modeling, as they had measurements at the first three time points only and modeling optimization had placed them into a class of their own. These exclusion criteria resulted in a final sample size of 1,132 children. Just under half of the children were female (49.1%). The majority of parents described themselves as white (81.1%), with 12.1% black or African American (black/AA), 1.5% Asian or Pacific Islander, less than 1% American Indian, Eskimo, or Aleutian, and 4.9% designated as Other. Parents of 5.9% of the children selected Hispanic ethnicity.

TABLE 1.

Breakdown of demographic characteristics by weight predictors in the final sample (N = 1,132)

	Birth weight (hg) for GA, mean (SD)	Weight increase (%) at 0–15 months, mean (SD)
Full sample	35.2 (4.6)	210.2 (42.5)
Race
Black/African American	33.6 (3.8)	228.0 (48.6)
Other races	35.5 (4.6)***	208.8 (41.0)***
Ethnicity
Hispanic	35.7 (5.3)	207.2 (44.8)
Non-Hispanic	35.2 (4.5)	210.4 (42.3)
Gender
Female	34.6 (4.6)	206.4 (43.8)
Male	35.9 (4.5)***	213.8 (40.9)**

Only black/African American differed from other races on predictors; therefore, the other races were combined.

Asterisks indicate significant mean difference (Welch’s t test) from the category directly above:

^**P < 0.01;

^***P < 0.001.

GA, gestational age; hg, hectogram; SD, standard deviation.

We confirmed that the SECCYD data set showed the same shift in children’s weight since the 1970s, when the bulk of the measurements underlying the CDC’s growth charts were taken, as has been observed in previous longitudinal data sets. The median (50^th percentile) SECCYD BMI percentile fell at the CDC’s 69th percentile. Seventeen percent of the SECCYD population exceeded the CDC’s 95th percentile cutoff for obesity at eighth grade, triple the number of children with obesity in the 1970s when the CDC created its weight benchmarks.

Latent growth modeling

All models used to identify classes were unconditional; that is, they did not include covariates or predictors of membership, only BMI percentiles at the nine time points. The best LCGA solution yielded five classes, including two that were generally low and stable across time, one that started in the upper third of BMI and rose still higher in early childhood and one that rose steadily from low to high through middle childhood (Figure 1A). The GMM solution yielded fewer classes, as expected (Figure 1B). GMM effectively combined the two low-stable classes seen in the LCGA solution into one and omitted the low-to-high rising trajectory altogether.

Figure 1.

BMI trajectories, 24 months to eighth grade, estimated by (A) LCGA and (B) GMM among children born at full term (37–41 weeks). [Color figure can be viewed at wileyonlinelibrary.com]

For the final trajectory solution, we retained the low-to-high trajectory estimated by LCGA, as this pattern resulted in high BMI and was of particular developmental interest. We defined the parameters of the low-to-high class in GMM, which then classified approximately 12% of the sample into that class and the remainder into the other three. The final four BMI trajectories (% sample) from 2 years to eighth grade were low-stable (28%), low-to-high rising (12%), median-stable (29%), and high-rising (31%) (Figure 2). The distribution of the low-stable class was noticeably bimodal, which was expected, given that it effectively combined two low-stable classes found by LCGA. Full LCGA, GMM, and final model diagnostics are found in Supporting Information Tables S1 and S2.

Figure 2.

Final BMI trajectories, 24 months to eighth grade, among children born at full term (37–41 weeks). For the final trajectory solution, parameters of the low-to-high class found in the LCGA solution (Figure 1A) were specified explicitly in a GMM, and the remaining classes were estimated freely. [Color figure can be viewed at wileyonlinelibrary.com]

As seen in Figure 2, the four trajectories start to be visually distinguishable at around 36 months, with the highest BMI percentiles observed in fifth and sixth grades. Children in the high-rising trajectory retained the highest BMI throughout, and 63% exceeded the cutoff defining obesity at least once during childhood (Table 2). The low-to-high rising trajectory, visually distinguishable from the low-stable trajectory by 54 months and rising steeply throughout middle childhood, resulted in the next highest BMI at eighth grade. Almost 70% of children in the median-stable trajectory were classified as overweight (≥85th percentile BMI) at least once, reflecting the population shift upward. The ceiling at 100th percentile BMI affected the distribution of data in the high-rising class, which had noticeable left skewness. This feature of the data could only bias hypothesis testing conservatively, that is, reduce chances of finding differentiation between the high-rising and other trajectories.

TABLE 2.

BMI trajectories: Demographic and overall weight statistics by class

Trajectory (% of sample)		Mean BMI percentile	Female (%)	Black/AA (%)	Mean BWT for GA (hg)	Mean % infant WTINC (0–15 mo)	% Overweight	% Obesity
High-rising	(31%)	86th	46.4	15.8	35.99	218.1	91	63
Median-stable	(29%)	68th	47.9	12.1	35.84	212.0	69	21
Low-to-high	(12%)	55th	44.4	13.9	34.63	201.9	44	14
Low-stable	(28%)	32nd	54.6	7.4	34.03	202.6	6	1

Measurements for mean BMI come from nine waves.

% overweight represents percent of children who had BMI percentile ≥85th at one or more time points; % obesity represents percent of children who had BMI percentile ≥95th at one or more time points.

Black/AA, black/African American; BWT for GA, birth weight for gestational age; WTINC, weight increase; hg, hectogram.

Predicting trajectory membership

In both between-race t tests and predictions of class membership (not shown), the only contrast of statistical significance was vis-à-vis the black/AA classification; thus, we coded one variable as black/AA = 1 and all others combined = 0. Hispanic ethnicity was omitted from further analyses, as it was unrelated to birth weight for gestational age, percent weight increase at 0 to 15 months, or membership in the higher BMI trajectory classes. The final variables tested for their combined effect on trajectory membership were the black/AA variable, gender, birth weight for gestational age, and percent weight increase at 0 to 15 months. We tested these together in multinomial logistic regression to develop risk algorithms for following higher BMI trajectories.

Risk ratios (RRs) and prediction formulas from the regression model are in Table 3. The black/AA variable predicted greater risk for membership in all higher BMI classes relative to the low-stable class. Girls were more likely to follow the high-rising and median-stable weight trajectories than were boys, when controlling for the other covariates; boys were more likely than girls to follow the low-to-high trajectory.

TABLE 3.

Relative risk associated with each predictor for membership in each of the higher BMI trajectories vs. the low-stable trajectory, with the related formula for an individual’s predicted risk of membership

BMI trajectory	RR	(95% CI)	Formula for total risk of membership vs. membership in the low-stable trajectory
High-rising trajectory			$$\begin{gathered} RR=\exp (-18.21+0.29\left[\begin{array}{c}1=F\\ 0=M\end{array}\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.97\left[\begin{array}{c}1=Black/AA\\ 0=All\hspace{0.17em}others\end{array}\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.32\left[BWT\hspace{0.17em}for\hspace{0.17em}GA\hspace{0.17em}\left(hg\right)\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.03\left[\%\hspace{0.17em}WTINC\right]) \end{gathered}$$
Female (1) vs. male (0)	1.33	(1.13–1.57)
Black/AA (1) vs. all others (0)	2.63	(2.00–3.45)
BWT for GA (hg)	1.38	(1.35–1.41)
Infant weight increase (%)	1.03	(1.03–1.04)
Median-stable trajectory			$$\begin{gathered} RR=\exp (-14.47+0.22\left[\begin{array}{c}1=F\\ 0=M\end{array}\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.66\left[\begin{array}{c}1=Black/AA\\ 0=All\hspace{0.17em}others\end{array}\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.26\left[BWT\hspace{0.17em}for\hspace{0.17em}GA\hspace{0.17em}\left(hg\right)\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.03\left[\%\hspace{0.17em}WTINC\right]) \end{gathered}$$
Female (1) vs. male (0)	1.24	(1.05–1.46)
Black/AA (1) vs. all others (0)	1.95	(1.47–2.59)
BWT for GA (hg)	1.30	(1.27–1.32)
Infant weight increase (%)	1.03	(1.02–1.03)
Low-to-high trajectory			$$\begin{gathered} RR=\exp (-3.03-0.35\left[\begin{array}{c}1=F\\ 0=M\end{array}\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.76\left[\begin{array}{c}1=Black/AA\\ 0=All\hspace{0.17em}others\end{array}\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{0.17em}+0.05\left[BWT\hspace{0.17em}for\hspace{0.17em}GA\hspace{0.17em}\left(hg\right)\right]) \end{gathered}$$
Female (1) vs. male (0)	0.70	(0.56–0.88)
Black/AA (1) vs. all others (0)	2.13	(1.50–3.02)
BWT for GA (hg)	1.05	(1.02–1.07)
Infant weight increase (%)	1.00	(1.00–1.00)

Hispanic ethnicity not retained as significant predictor as assessed either by the P value of its coefficient or by a χ² difference test between this model and the larger one including ethnicity.

Black/AA, black/African American; RR, risk ratio for membership in tested trajectory vs. low-stable trajectory; BWT for GA, birth weight for gestational age, with RR interpreted in terms of a 1-hg increase; WTINC, percent infant weight increase from 0 to 15 months; hg, hectogram.

Both higher birth weight for gestational age and greater percent weight increase at 0 to 15 months predicted greater risk for following the high-rising and the median-stable trajectories versus the low-stable trajectory. Birth weight did not moderate the effect of rapid weight gain; that is, greater percent weight increase predicted future membership in the high-rising and the median-stable BMI trajectories regardless of the child’s starting weight. The low-to-high rising trajectory was differentiated the least by the tested predictors: black/AA males were more likely to be members, but the early weight variables were only weakly related.

The formulas in Table 3 may be used to calculate RRs and illustrate the important effects of the early weight variables. For example, a girl coded as black/AA with approximately average birth weight and average weight increase from 0 to 15 months has an RR of exp(−18.21 + 0.29[1] + 0.97[1] + 0.32[34 hectograms] + 0.03[206% weight increase]) ≈1.10. That is, at 15 months old, she is only 10% more likely to follow the high-rising trajectory than the low-stable. But if she has gained approximately 1 SD greater percent weight increase, the RR increases to exp(−18.21 + 0.29 [1] + 0.97[1] + 0.32 [34 hectograms] + 0.03 [250% weight increase]) ≈ 4.18. Even at average birth weight, a 1-SD increase in percent weight gain at 15 months makes her membership in the high-rising versus low-stable class four times more likely. Similarly, a 1-SD greater percent weight gain at 15 months increases the RR of a non-black/AA boy with average birth weight from ≈0.76 to ≈2.59.

The high-rising and median-stable trajectories had similar BMI starting points at 24 months but markedly different BMI outcomes. A post hoc logistic regression predicted membership in the high-rising versus the median-stable trajectory only by birth weight for gestational age and percent weight increase at 0 to 15 months (RR = 1.01 and 1.05, respectively). At the mean birth weight (35.2 hectograms) and 1-SD greater percent weight increase, a child was 41% more likely to follow the high-rising rather than median-stable trajectory. At 2-SD greater percent weight increase, the likelihood rose to 88%. Neither the black/AA variable nor gender influenced the differential likelihood at 15 months old between these two trajectories, which together comprised 60% of the study population.

Discussion

Childhood overweight and obesity have been linked to a range of negative outcomes for later health, including heart disease and resulting premature death (1,2). Predicting who has the highest risk for high BMI in late childhood and early adolescence is thus an important task for both clinicians and public health practitioners who want to reduce overweight and obesity. In this study, we used prospective longitudinal data to identify patterns of weight gain and then identified early-life predictors of those patterns. Because the predictors we used are easily available to physicians, they allow physicians to quickly identify children at highest risk for later overweight and obesity and to focus their educational and monitoring efforts on these children and their families.

Using data from 11 waves of a prospective longitudinal study, we identified four classes of change in BMI percentiles over the period from 24 months to roughly age 13, with two classes (high-rising class and low-to-high class) exhibiting steep increases in BMI that resulted in a risk for overweight and obesity by age 13. We also demonstrated the relative importance of the demographic and weight predictors, in that risk for membership in the high-BMI trajectories can be determined as early as 15 months old. As birth weight for age differentiated all higher BMI trajectories from the low-stable class, high birth weight should trigger parents’ and physicians’ attention to proper nutrition in the first few months and years of life. However, birth weight only weakly distinguished the rising classes (high-rising and low-to-high) from their stable counterparts (median-stable and low-stable). Of greater clinical significance were the findings with respect to percent increase in weight from 0 to 15 months. Rapid weight gain in infancy predicted a high-rising trajectory independent of birth weight, indicating no positive health effect of catch-up weight gain in low-birth-weight full-term babies, at least not with respect to BMI. That is, rapid weight gain appears equally deleterious for both low- and high-birth-weight babies.

This finding both complicates and informs interpretation for babies identified as black/AA, as their birth weight was significantly lower and their weight increase from 0 to 15 months significantly greater than their peers. Whether parents had self-identified as black/AA predicted membership in all trajectories versus the low-stable, controlling for the other covariates. Thus, race is a marker for risk over and above the tested biological variables, although how race is related to biological differences is unclear (17). Important covariates and/or mediators related to differences in racial identification and weight outcomes are certainly missing from this model, such as maternal weight gain and obesity (23), postnatal nutrition and behavioral choices (24), and socioeconomic status (25). The race variable did not differentiate the high-rising from the median-stable trajectory with a much lower BMI at adolescence; early rapid weight gain was the primary distinguishing variable between those two. Our findings should not be read to imply that black families have an inherent risk for obesity. Rather, it is more likely that black race is a marker for risks linked with social contexts that make low birth weight, early rapid weight gain, and obesity more likely (e.g., poverty, stress, food deserts, violent neighborhoods) and that are experienced more by black families than other families because of factors such as racism, discrimination, and unfair public policies and distribution of public resources.

The multivariate model also clarified the role of gender in predicting weight trajectories. Prevalence reports have tended to suggest girls are less likely than boys to develop obesity (26), or they have omitted gender as a predictor altogether (2,7). Here we found girls more likely than boys to be in the higher BMI classes versus the low-stable when controlling for race, birth weight, and percent infant weight increase. We did not explore differences within gender and across ancestry groups, as has been possible elsewhere (27), because of small sample sizes in such subgroups; nor did we have reliable measures of BMI versus adiposity available. Further research on the timing of sexual maturity and the health outcomes of boys and girls in these trajectories is needed to understand real risk.

The main limitation of this study is that the sample was not nationally representative and thus cannot provide direct comparisons to CDC growth charts. We judge that limitation to be outweighed by the reliable, technician-measured height and weight data in this study and the longitudinal research design. Although our infant weight gain measure that extends to 15 months is longer than some previous studies that have looked only at the first few months, we think it is important to examine early weight gain across the window in which children are increasing in mobility (both crawling and walking) because children who experience significant weight gain despite this rapid increase in mobility (i.e., exercise) are likely to be different metabolically or to have different home environments than children who do not.

The clinical implications of the heterogeneous weight change trajectories we have identified are significant. The CDC’s BMI charts for children were developed on cross-sectional data collected from 1963 to 1994 (28). Thus, although they are used by clinicians and parents to track past and predict future BMI growth in children, the charts were not actually derived from data on change in BMI over time. By characterizing within-child growth patterns across childhood, the growth curves presented in this study are a major improvement upon the CDC growth charts. Distilling common patterns of change in BMI percentile offers opportunities in this and other data sets to explore complex relations among other known covariates and weight change, such as breastfeeding history, maternal BMI, and pubertal timing.

Conclusion

At 15 months of age, birth weight for gestational age and percent weight increase are the strongest predictors for following a high-rising BMI trajectory of weight change across childhood. The exceptional low-to-high rising trajectory emerged later in childhood and was more likely to contain boys and children identified as African American. Identifying and intervening with families and children at risk for high and rapidly rising weight trajectories should be the focus of obesity prevention efforts for the highest value to public health.