The Early Childhood Longitudinal Study – Kindergarten Cohort (ECLS-K) is conducted by the National Center for Education Statistics using a nationally representative cohort of children who entered kindergarten during 1998-1999. The sample was drawn from approximately 1,000 schools, including public and private schools and full-day and part-day programs (17,18). Sampling was based on a multistage probability sampling design in which counties were selected, then schools, and finally children within schools. The ECLS-K collected information from children, parents, teachers, and schools by using a variety of formats, including face-to-face assessments or interviews, telephone interviews, and questionnaires (26,27). Trained assessors at the children's schools conducted the direct child assessment. Height and weight were measured in kindergarten, first grade, third grade, and fifth grade.

We used the ECLS-K Longitudinal Public-Use Data File, which combined data from kindergarten, first grade, third grade, and fifth grade, and included 17,565 children at baseline. We constructed the analytic sample as follows. First, we identified children in each round who were from either 2-parent or single-mother families (n = 14,831 for kindergarten, 13,123 for third grade, and 10,747 for fifth grade). Next, we included only children who had complete child direct assessment data (ie, measured height and weight) from kindergarten (n = 14,493), third grade (n = 11,855), and fifth grade (n = 10,036). Last, we included only observations with positive person-level sample weights.

The study outcomes were 1) BMI as a continuous variable, 2) obesity as a binary variable, and 3) change in BMI from kindergarten to fifth grade as a continuous variable. We used BMI to categorize children's weight status. BMI is calculated by dividing weight in kilograms by the square of height in meters (kg/m²) and is widely considered one of the best clinical weight criterion for children because BMI in childhood tracks well into adulthood for developing obesity (13-15). Percentile comparisons are based on the sex-specific BMI-for-age growth charts from the Centers for Disease Control and Prevention (CDC) (19,20).

We used multivariate regression to assess the effect of family structure on the study outcomes while controlling for sociodemographic variables that may influence these outcomes. We analyzed continuous outcomes by using linear regression and binary outcomes by using logistic regression.

The main explanatory variables in the regression models were measures of family structure: an indicator variable for living in a single-mother family (vs a 2-parent family) and indicator variables for the number of siblings, categorized as none (the omitted category), 1, 2, or 3 or more. We defined 2-parent families as families in which both the father and the mother (biological, adopted, or step) were living with the child in the household. Correspondingly, we defined single-mother families as families in which the child's mother was living with the child but the father was absent. We excluded single-father families because of their small number.

The covariates in the models included the following sociodemographic characteristics: indicator variables for the child's age; sex; race/ethnicity, categorized as non-Hispanic white (the omitted category), non-Hispanic black, Hispanic, or other; family income, categorized as poor (<1.00 times the federal poverty level — omitted category), low-income (1.00-1.99 times poverty), middle-income (2.00-3.99 times poverty), or high-income (more than 4 times the federal poverty level); mother's education, categorized as high school or less (omitted), high school graduate, some college, or bachelor's degree or higher; and mother's age (at time of child's enrollment in the study), categorized as 24 years or younger (omitted), 25 to 34 years, 35 to 44 years, or 45 years or older. We used the other covariates to capture the child's health at birth: premature birth of more than 2 weeks; birth weight, categorized as less than 2,000 g (omitted), 2,000 to 2,499 g, 2,500 to 2,999 g, 3,000 to 3,499 g, and 3,500 g or more; birth order, categorized as first (omitted), second, or third or higher. We combined birth weight less than 1,500 g with birth weight less than 2,000 g because of the small sample size of very low birth weight children in the ECLS-K cohort (0.9% of the sample had birth weight <1,500 g).

We weighted all analyses by using sample weights provided by ECLS-K that reflect the sample design and survey nonresponse, and we adjusted all standard errors for clustering by using the Huber-White sandwich estimator (21,22). We conducted all analyses using Stata version 10 (StataCorp LP, College Station, Texas). This study was exempt from institutional review because it used public-use data with no identifiable information.

To facilitate interpretation of the regression results, we used the method of recycled predictions to obtain the predicted mean values of the study outcomes for each type of family while adjusting for the covariates (6,23). Specifically, we used the estimated coefficients from the regression models to predict each outcome for each child, alternately assigning the child to each category of the family structure variable of interest (eg, single-mother vs 2-parent family) but leaving all other explanatory variables at their original values. Next, we averaged the predictions across all the children in the sample. This procedure yielded what the mean value of each outcome would be if all children in the sample lived in each particular type of family (eg, single-mother or 2-parent family) but otherwise retained the original values of all their other characteristics.

At enrollment in the study (ie, in kindergarten), excluding single-father and other family types, approximately one-fourth of the children were from single-mother families (Table 1). Nearly half of the children were girls, and most of them had at least 1 sibling. Approximately 60% of the children were non-Hispanic white, 15% were non-Hispanic black, and 18% were Hispanic.

Among all children with measured height and weight in kindergarten, mean BMI was 16.3 kg/m², and 13.0% of the children were obese. In third grade, mean BMI was 18.6 kg/m², and nearly 21% of the children were obese. In fifth grade, mean BMI was 20.6 kg/m², and 24% of the children were obese (data not shown).

Family structure was significantly associated with the obesity rate (Table 2). In each grade, children from single-mother families had higher rates of obesity than children from 2-parent families. Thus, in kindergarten, 14% of children from single-mother families were obese, compared with 13% of children from 2-parent families (P = .05). Similarly, in third grade, 23% of children from single-mother families were obese, compared with 20% of children from 2-parent families (P = .03). In fifth grade, 28% of children from single-mother families were obese, compared with 22% of children from 2-parent families (P = .003). Moreover, in each grade the number of siblings was negatively associated with the rate of obesity. We also observed similar associations between family structure and BMI.

We obtained adjusted results for BMI and risk of obesity by using the multivariate models and recycled prediction (Table 3). The number of parents in the household was not associated with BMI or risk of obesity in kindergarten or third grade, but by fifth grade children from single-mother families (26%, P = .05) were more likely to be obese than their peers from 2-parent families (22%). In every grade we found that children with no siblings had higher BMI and a higher probability of being obese than children with siblings (Table 3).

Children living with no siblings had a larger increase in BMI than children living with siblings. The increase in BMI was 4.7 for children with no siblings; 4.2 for children with 1 sibling (P = .02); 3.8 for children with 2 siblings (P = .01); and 3.7 for children with 3 or more siblings (P = .006) (data not shown). However, the increase in BMI from kindergarten to fifth grade did not differ significantly by number of parents in the household.

In kindergarten, more black (14%, P = .02) and Hispanic (17%, P < .001) children were obese than were white children (12%), and more poor children were obese (15%) than were children from high-income families (10%, P < .001). Similarly, in third grade, more black (23%, P = .02) and Hispanic (26%, P < .001) children were obese than white children (18%), and more poor children (24%, P < .001) were obese than were children in high-income families (16%). Mother's education also became a significant predictor of obesity in third grade (25% for high school or less, compared with 17% for college degree or higher, P = .03). Results for fifth grade were similar to those for third grade.

We repeated our analyses with indicator variables for the age spacing of the closest sibling, categorized as less than 1 year (omitted category), 1 to 2 years, 2 to 3 years, or 3 or more years. We found that age spacing between siblings was not associated with BMI or risk of obesity, and accounting for age spacing did not change our other findings.

We also examined several variables, including time spent watching television (in fifth grade), fast-food consumption (in fifth grade), exercise (in fifth grade), and child care before kindergarten, to assess whether they mediated or moderated the associations of family structure with BMI and risk of obesity that we found. Among these variables, only weekly television hours (measured as the sum of hours of television, videotapes, or DVDs watched outside of school) exhibited a small mediation effect for the association of single-mother family with risk of obesity. We found no other mediation effects and no moderation effects.