Period-specific growth, overweight and modification by breastfeeding in the GINI and LISA birth cohorts up to age 6 years

Abstract

Children’s weight/growth development is age-specific and may be influenced by breastfeeding. We therefore assessed velocities of weight, length, body-mass-index and overweight/obesity development from birth up to age 6 years overall and in relation to breastfeeding. The method of this study is based on pooled data of the birth-cohorts GINI-plus and LISA-plus and follows 7,643 healthy full-term neonates in four study-centers in Germany. Up to nine anthropometric measurements are available. Overweight/obesity is percentile-defined according to WHO-Child-Growth-Standards. Fully-breastfed is defined as breastfed for at least 4 months. Piecewise-linear-random-coefficient-models were applied to assess growth trajectories and velocities between 0–3, 3–6, 6–12, 12–24 and beyond 24th months. Velocities for weight-, length- and BMI-development are highest in the first 3 months after birth and diminish, with differing pace, in the periods that follow. For overweight and obesity, peak-velocities are estimated in periods 6–12 and 3–6 months. The difference in the velocity of weight gain for breastfed vs. other children is −18 g/month in the first 3 month, −93 g/month between month 3 and 6, −14 g/month between month 6 and 12 and −3 g/month beyond the 24th month. Velocities in length are not different between breastfed and non-breastfed children. Over time, a slightly lower risk (difference < 2%) of being overweight was estimated for breastfed children, after adjustment for study-center, socio-economic-status and maternal smoking in pregnancy. Infants fully-breastfed gain less weight, but grow equally in length in the first 12 months of life versus mixed or formula-fed children. The protective effect of breastfeeding on becoming overweight is related to its weight-velocity-modifying-effect in early infancy.

Appendices

Appendix

Detailed information on statistical analysis

Piecewise linear random coefficient models were applied to assess growth trajectories and velocities between 0–3, 3–6, 6–12, 12–24 months and beyond the 24th month. These models allow the longitudinal data structure to be accounted for by including subject specific random effects and a nonlinear age effect can be modeled by the piecewise linear functions (polynomial splines). Such longitudinal models are described in detail in the books of Singer et al. and Fitzmaurice et al. [18, 19]. We used four knots at 3, 6, 12 and 24 months to connect the slopes of the five time segments. The choice of the knots was based on the literature, in which different time windows for rapid weight gain are discussed [23–26]. To account for the known sex specific difference in birth weight and length we included a main effect for sex in each model.

Formally, the basic piecewise linear random coefficient model at hand (Model A) can be expressed as follows:

$$ \begin{aligned} Y_{ij} & = \beta_{0i} + \beta_{1i} {\text{Age}}_{ij} + \beta_{2i} ({\text{Age}}_{ij} - 3)_{ + } + \beta_{3i} ({\text{Age}}_{ij} - 6)_{ + } \\ & \quad + \beta_{4i} ({\text{Age}}_{ij} - 12)_{ + } + \beta_{5i} ({\text{Age}}_{ij} - 24)_{ + } + \beta_{6} {\text{boy}}_{i} + e_{ij} = \eta_{ijA} + e_{ij} \\ \end{aligned} $$

(1)

where Y _ij is the respective continuous outcome (i.e. length, weight or BMI) for child i at measurement j and Age _ij is age since birth, coded in months, for each child i at measurement j (calculated from the exact age in days). The term (Age _ij  − c)₊ with knots c ∈ {3, 6, 12, 24} is equal to (Age _ij  − c) if Age _ij  > c and equal to 0 if Age _ij  < c. The effects β _ki for k = 0,…,5 consist each of a population averaged fixed effect β _k and a subject specific random effect u _ki , as given by: $\beta_{0i} = \beta_{0} + u_{0i}, \beta_{1i} = \beta_{1} + u_{1i}, \beta_{ 2i} = \beta_{ 2} + u_{ 2i}, \beta_{ 3i} = \beta_{3} + u_{3i}, \beta_{4i} = \beta_{4} + u_{4i}$ and $\beta_{5i} = \beta_{5} + u_{5i}$. Hence, a subject specific intercept u _0i as well as five subject specific slopes u _1i ,…, u _5i are estimated. The subject specific random effects vectors u _i  = (u _0i ,…,u _5i )^T are assumed to be mutually independent for all i and normally distributed with zero mean and a covariance matrix Σ, i.e. u _i  ~ N (0, Σ). The diagonal of Σ contains the coefficient specific variances σ ² _k for k = 0,…,5. The error terms e _ij are also assumed to be normally distributed and identical and mutually independent for all i, j, i.e. e _ij  ~ N (0, σ ² _e ) i.i.d. In addition, they are considered as independent from the random effects. The short notation η _ijA in [1] stands for the predictor of Model A and is introduced by reason of comparability between the different models.

As for interpretation for the regression coefficients, ß ₁ can be regarded as the population baseline velocity of change for the respective outcome and hence, u _1i is the subject specific deviation from this population baseline. The term (Age _ij  − 3)₊ represents the time since the age of 3 months until measurement j of child i, consequently ß ₂ represents the population based deviation from the slope ß ₁ in the following time period and u _2i stands for the associated individual deviation. For all other knots the coding and interpretation is analog. Thus, each child can have his own baseline value at birth and a child specific slope or linear trajectory in each time period, which yields to a subject specific non-linear growth pattern by the cumulative combination of the several linear growth estimates. The growth rate GR at the age period k ∈ {0–3, 3–6, 6–12, 12–24, 24–72} months is thus the cumulative period specific sum of the estimated regression coefficients, for the first three periods it can be expressed as follows: ${\text{GR(0}} - 3 )= 3 \times \beta_{1} ,{\text{GR(0}} - 6 )= 6 \times \beta_{1} + 3 \times \beta_{2} ,{\text{GR(6}} - 12 )= 12 \times \beta_{1} + 9 \times \beta_{2} + 6 \times \beta_{3} .$

To ease interpretation and to spare the reader the trouble of calculation we do not report the single slope coefficients in the result section but we do report the calculated absolute growth rates (velocities) of the outcome per month in the respective time period (Tables 2, 4). The subject specific variation of the intercept terms (initial status at birth) and of the period specific growth rates (rate of change per month in period) are expressed as 95%-reference ranges and listed in Table 3. A reference range is the range in which 95% of the estimated subject-specific intercepts or slopes (here for the calculated growth rates) are located, formally: $\hat{\beta }_{k}$ ± 1.96 × estimated standard deviation of the subject specific estimates $\hat{u}_{ik}$ (square root of the estimated random effect variance $\hat{\sigma }_{k}^{2}$). Note that if the growth rate is a combination of several slopes (e.g. for period 3–6 months, which is calculated as the sum of the slopes in period 0–3 and 3–6), then the standard deviation is calculated as the square root of the sum of the respective variances of the slopes and the sum of two times the respective covariances of these random effects.

To what extent individual initial status of the outcome at birth and individual change rates co-vary between the different time windows is expressed as correlations (derived from the estimated random effects covariance matrix $\hat{\Upsigma }$) and is listed in the lower part of Table 3. We report these subject specific variations of growth rates only for weight and length because for the models regarding BMI, overweight and obesity development no reliable random variation in growth rates (beyond the intercept term) could be estimated.

For the dichotomous outcomes of overweight and obesity generalized random coefficient models with logit-link function were applied. Therefore the outcome Y _ij was assumed to follow a binomial distribution with probability π _ij , i.e. Y _ij  ~ B (1, π _ij ). Hence, the model (Model A) can be expressed as follows:

$$ \pi_{ij} = E(Y_{ij} /\eta_{{ij{\text{A}}}} ) = {\frac{1}{{1 + \exp ( - \eta_{{ij{\text{A}}}} )}}} $$

(2)

where Y _ij is a dichotomous outcome (i.e. overweight or obesity) and η _ijA is the predictor as in [1]. Since they are easier to interpret, we report probabilities in Table 4 of the result section (instead of using logarithmic odds).

In Tables 2 and 3 of the result section there are three models for each outcome presented. The respective Model A has already been introduced in [1] and [2], depending on the outcome. It gives estimates for the baseline value (initial status) of the outcome and the time period specific velocities (absolute change of the outcome per month in the respective period) for the five time segments with sex as the only covariate. Model B aims at investigating the influence of breast feeding on the rates of change for the five time periods by including a main effect for breastfeeding (BF) as well as interaction effects with the piecewise linear terms. Formally, Model B can be obtained by replacing the predictor η _ijA of Model A in Eqs. 1 and 2 by η _ijB, as given by:

$$ \begin{aligned} \eta_{ijB} & = \eta_{{ij{\text{A}}}} + \beta_{7i} {\text{Age}}_{ij} \times {\text{BF}}_{i} + \beta_{8i} ({\text{Age}}_{ij} - 3)_{ + } \times {\text{BF}}_{i} \\ & \quad + \beta_{9i} ({\text{Age}}_{ij} - 6)_{ + } \times {\text{BF}}_{i} + \beta_{10i} ({\text{Age}}_{ij} - 12)_{ + } \\ &\quad \times {\text{BF}}_{i} + \beta_{11i} ({\text{Age}}_{ij} - 24)_{ + } \times {\text{BF}}_{i} + \beta_{12} {\text{BF}}_{i} . \\ \end{aligned} $$

(3)

In the result section we do not show the single interaction estimates but report the absolute growth rates for the breastfed and for the other children in two different columns. These interaction effects allow an evaluation as to whether breastfeeding influences the velocities of growth in the different time windows.

Model C accounts for the potential confounding effects of maternal smoking in pregnancy (Smoke), study center (Center) and socio-economic status (SES) in addition to the breastfeeding Model B by adjusting for the respective number of dummy coded categorical variables. Hence, the predictor for Model C can be expressed as follows:

$$ \eta_{{ij{\text{C}}}} = \eta_{{ij{\text{B}}}} + {\text{Smoke}}_{i} + {\text{SES}}_{i} + {\text{Center}}_{i} . $$

(4)

Descriptive analyses were conducted by the statistical software SAS, version 9.1.3 [36]. All longitudinal analyses were performed with the special purpose software for multilevel modeling MLwiN, version 2.02 [37]. (Tables 5, 6, 7 and 8).

Table 5 Detailed listing of regression coefficients (fixed effects) from which change rates of Table 2 were derived

Table 6 Random effects of models listed in Table 5

Table 7 Detailed listing of logistic regression coefficients (fixed effects) from which change rates of Table 4 were derived

Table 8 Random effects of models listed in Table 7

GINI-plus study group

Institute of Epidemiology, Helmholtz Zentrum Muenchen-German Research Center for Environmental Health, Neuherberg (Wichmann HE, Heinrich J, Schoetzau A, Popescu M, Mosetter M, Schindler J, Franke K, Laubereau B, Sausenthaler S, Thaqi A, Zirngibl A, Zutavern A, Filipiak B, Gehring U); Department of Pediatrics, Marien-Hospital, Wesel (Berdel D, von Berg A, Albrecht B, Baumgart A, Bollrath C, Büttner S, Diekamp S, Groß I, Jakob T, Klemke K, Kurpiun S, Möllemann M, Neusüss J, Varhelyi A, Zorn C); Ludwig Maximilians University of Munich, Dr. von Hauner Children’s Hospital (Koletzko S, Reinhard D, Weigand H, Antonie I, Bäumler-Merl B, Tasch C, Göhlert R, Sönnichsen C); Clinic and Polyclinic for Child and Adolescent Medicine, University Hospital rechts der Isar of the Technical University Munich (Bauer CP, Grübl A, Bartels P, Brockow I, Hoffmann U, Lötzbeyer F, Mayrl R, Negele K, Schill E-M, Wolf B); IUF-Environmental Health Research Institute, Düsseldorf (Krämer U, Link E, Sugiri D, Ranft U).

LISA-plus study group

Institute of Epidemiology, Helmholtz Zentrum Muenchen-German Research Center for Environmental Health, Neuherberg (Wichmann HE, Heinrich J, Bolte G, Belcredi P, Jacob B, Schoetzau A, Mosetter M, Schindler J, Höhnke A, Franke K, Laubereau B, Sausenthaler S, Thaqi A, Zirngibl A, Zutavern A); Department of Pediatrics, University of Leipzig (Borte M, Schulz R, Sierig G, Mirow K, Gebauer C, Schulze B, Hainich J); Institute for Clinical Immunology and Transfusion Medicine (Sack U, Emmrich F); Department of Pediatrics, Marien-Hospital, Wesel (von Berg A, Schaaf B, Scholten C, Bollrath C, Groß I, Möllemann M); Department of Human Exposure-Research and Epidemiology, UFZ-Center for Environmental Research Leipzig-Halle (Herbarth O, Diez U, Rehwagen M, Schlink U, Franck U, Jorks A, Röder S); Department of Environmental Immunology, UFZ-center for Environmental Research Leipzig-Halle (Lehmann I, Herberth G, Daegelmann C); Ludwig Maximilians University Munich, Dr. von Hauner Children’s Hospital, Department of Infectious Diseases and Immunology (Weiss M, Albert M); Friedrich-Schiller-University Jena, Institute for Clinical Immunology (Fahlbusch B), Institute for Social, Occupational and Environmental Medicine (Bischof W, Koch A); IUF-Environmental Health Research Institute, Düsseldorf (Krämer U, Link E, Ranft U, Schins R); Clinic and Polyclinic for Child and Adolescent Medicine, University Hospital Rechts der Isar of the Technical University Munich (Bauer CP, Brockow I, Grübl A); Department of Dermatology and Allergy Biederstein, Technical University Munich (Ring J, Grosch J, Weidinger S).