Introduction. R Hardy 1 *, M Wadsworth 1 and D Kuh 1. University College Medical School, London, UK

Transcription

1 (2000) 24, 725±734 ß 2000 Macmillan Publishers Ltd All rights reserved 0307±0565/00 $ The in uence of childhood weight and socioeconomic status on change in adult body mass index in a British national birth cohort R Hardy 1 *, M Wadsworth 1 and D Kuh 1 1 Medical Research Council National Survey of Health and Development, Department of Epidemiology and Public Health, Royal Free & University College Medical School, London, UK OBJECTIVE: To investigate the effect of childhood weight and childhood socioeconomic status on the pattern of change in body mass index (BMI) between 20 and 43 years. METHODS: A British birth cohort study where the survey members have been followed up regularly since their birth in 1946, with the most recent of 19 follow-ups when the cohort were aged 43 years. BMI was available at 20, 26, 36 and 43 years of age and thus multilevel models for repeated outcome measures were used to model the patterns of change in BMI. RESULTS: The rate of increase in BMI with age was non-linear, with the rate of increase in mean BMI accelerating with increasing age at different rates for men and women. The mean BMI for men was higher than that for women at all ages. Childhood manual social class, de ned in terms of father's occupation, and high relative weight at 14 years of age were associated with higher mean BMI across adult life, and these effects increased with age. The effects of childhood relative weight and social class were independent of educational attainment and adult social class. CONCLUSION: The study provides evidence of a long-term effect of childhood social and biological circumstances on BMI. The pathways underlying these relationships may be social or biological, but are not yet fully understood. (2000) 24, 725±734 Keywords: body mass index; change; childhood; cohort study; socioeconomic status Introduction Overweight and obesity in middle life have been associated with cardiovascular disease, adult-onset diabetes and various types of cancer, 1±8 as well as with mortality. 9±13 It has also been observed that obesity in childhood is associated with a higher risk of early cardiovascular disease mortality in adulthood, 14 but the relative contributions of childhood and adult obesity on mortality are not clear. Obesity and overweight in childhood predispose to overweight and obesity in adulthood, 15,16 although the correlations between childhood and adult relative weight have tended to be weak. 17 In addition, low birth weight has been identi ed as a marker of biological programming of adult chronic disease 18,19 and birthweight has generally been positively correlated with adult body mass index (BMI). 20 However, an inverse relation between birthweight and waist ± *Correspondence: R Hardy, Medical Research Council National Survey of Health and Development, Department of Epidemiology and Public Health, Royal Free & University College Medical School, 1 ± 19 Torrington Place, LONDON WC1E 6BT, UK. rebecca.hardy@ucl.ac.uk Received 10 May 1999; revised 4 October and 28 November 1999; accepted 10 January 2000 hip ratio, a measure of central adiposity, has been observed. 21,22 Hence, weight throughout childhood plays a part in the development of overweight in adulthood and subsequent disease. Associations between childhood socioeconomic status and adult BMI and obesity which are independent of adult socioeconomic status have previously been reported, 17,23 ± 25 suggesting possible continuing in uences of the childhood environment into adulthood. The Medical Research Council (MRC) National Survey of Health and Development (NSHD) is a longitudinal birth cohort study in which heights and weights have been collected 4 times in adulthood; these were self-reported at 20 and 26, and measured at 36 and 43 years of age. It was therefore possible to investigate how BMI, a commonly used measure of adult relative weight from which overweight and obesity are derived, and its change with age between 20 and 43 years depend on childhood weight, including birthweight, and childhood socioeconomic status. Since the repeated measures of BMI on the same individual will be correlated, statistical methods which account for this data structure are required. The change in mean BMI with age may be modelled as a function of both time constant and time varying explanatory variables, using, for example, multilevel (or random effects) models for longitudinal data. 26,27

2 726 Methods The MRC National Survey of Health and Development is a socially strati ed cohort of 2548 women and 2814 men born in one week in 1946 in England, Scotland and Wales. There have been 19 contacts with the whole cohort from birth to the most recent home interview by research nurses when survey members were 43 years of age. 28,29 At this visit, 6.8% of the original cohort had died and 12.1% had permanently withdrawn from the study, 11.5% were living abroad and 8.8% had temporarily refused to participate or could not be traced. The population interviewed at the age of 43 years were, in most respects, representative of the native population of that age. 30 Height and weight Birthweight was recorded to the nearest quarter of a pound and converted to grams. Heights and weights were measured at ages 7, 11 and 14 y by school doctors or nurses. At ages 20 and 26 y, heights and weights were self-reported. The accuracy of these reported records have been investigated in a subsample of the population 15 and it was observed that an underestimate of the prevalence of overweight at these ages may have resulted. However, other research suggested that the magnitude of the bias involved was likely to be small. 31 At ages 36 and 43 y survey members were interviewed by trained research nurses in their own homes. At these interviews height, to the nearest 0.5cm, was measured using a portable stadiometer. Weight, to the nearest 0.5 kg below, in light indoor clothing was also measured, and was adjusted by subtracting 0.5 kg for women and 1 kg for men as a correction for the clothes worn. The outcome measure of weight for height in adulthood was expressed in terms of BMI, that is weight=height 2. An index of relative weight, de ned as body weight as a percentage of a standard weight calculated for speci c height, age and sex, 32,33 was used as the childhood measure of weight for height. Since the sample size of the NSHD was suf ciently large, standard weights were calculated from the total NSHD study sample. 15 Social class Childhood social class was de ned in terms of father's occupation when the survey member was 4 years old, and categorised into either manual or non-manual. Adult social class, again grouped into manual and non-manual, was considered as a time-varying explanatory variable and was taken as the social class, based on occupation of the head of the household, at ages 26, 36 and 43 y. If the current social class was not available, social class at the nearest previous age was used. At age 20 y, there was no social class information and so father's social class when the survey members were 15 y was used. The use of data from previous time points enabled the maximum possible sample size to be included in the analysis. Educational quali cations achieved by the age of 26 y were also considered as a measure of socioeconomic status, and survey members were divided into those with no quali cations, below ordinary secondary quali cation, ordinary secondary quali cations (`O' level or equivalent), advanced secondary quali cations (`A' level or equivalent) and higher education (degree or equivalent). Statistical methods The mean BMI, and the proportion overweight, were calculated at each age for different subgroups de ned by each of the childhood risk factors. Statistical modelling and testing was carried out using multilevel models tted using the computer package MLn. 34,35 As there were up to four outcome records for each individual, the data could be considered as hierarchical in structure. Hence, age (or measurement occasion) was de ned to be `level 1' and survey member as `level 2'. In general, y ij denotes the outcome of subject j ( j ˆ 1,..., n) on measurement occasion i (i ˆ 1,..., m i ) and x ij the time at which that measurement was taken. A basic multilevel model for repeated measures over time can be written as y ij ˆ a bx ij u j v j x ij e ij 1 where the xed parameter a represents the mean intercept and the xed parameter b the mean slope. The parameters u j and v j are the random (betweenindividual) effects which allow each individual to have their own intercept and slope respectively. Thus, u j indicates the deviation of the jth individual's intercept from the overall mean intercept and, similarly, v j the deviation from the overall slope. The variance of these random effects parameters may be estimated as well as the covariance between them. Similarly, the variance of the within-individual variation (e ij ) is estimated. Other explanatory variables, which may be constant or may vary over time, can be added as xed effects, while both the within-individual (`level 1') and between-individual (`level 2') variation can be modelled in relation to additional factors. In the current example, y ij was the BMI for survey member j at occasion i, where i took the values 1, 2, 3 and 4, and x ij was the age at which each measure of BMI was collected (20, 26, 36 and 43 y). The intercept represented mean BMI at age 20 y and the slope the mean linear change in BMI for each yearly increase in age. The statistical signi cance of each term added to the model was assessed (as in standard linear regression) using the test statistic obtained by dividing the regression coef cient by its standard error. 26 A test statistic greater than 1.96 (the standardised normal deviate exceeded with probability 0.05 under the null hypothesis) indicates that the effect is signi cant at

3 the 5% level. The model building process was initiated by tting a model containing only the variables representing age and sex. Birthweight, followed by the three measures of childhood relative weight, centred around 100 units (standard height, sex and age speci c weight), were added. Associations between these continuous risk factors and BMI were tested for deviations from linearity. Childhood social class was then considered in addition to relative weight. The regression coef cients relating to the time constant risk factors indicate the average effect on BMI over all ages. The interaction terms between each childhood factor remaining in the adjusted model and age were tted in order to test whether the rate of change in BMI varied across the different levels of the factor. In the presence of an interaction, the main effect could be interpreted as the difference in BMI at age 20 y and the differences at older ages could be calculated using the coef cients for both main effect and interaction. Interactions between childhood risk factors were also investigated. The nal childhood risk factor model was adjusted for adult social class and educational attainment in order to assess whether the in uence of the childhood factors, was independent of adult social factors. The pattern of both between-individual and within-individual random variation was also investigated. It was not necessary for each individual to have all outcome measures recorded, but in order to obtain unbiased estimates in the presence of missing data, it must be assumed that responses are missing at random (MAR); 36 that is, the probability of any BMI measure being missing may depend upon observed, but not unobserved, measures. The pattern of missing data was investigated. Finally, all analyses were repeated using the subgroup of cohort members with all four BMI measurements and results are reported where they differ from those of the main analysis. Exploratory analysis A variety of patterns of missing data as well as different growth patterns over time were observed across individuals in these data. The mean BMI, calculated separately for each measurement occasion, increased with age for both sexes, and at every age the mean BMI was lower among women than men. Among those overweight (relative weight 20% above standard) at 14 y, however, mean BMI was higher in women than men (Figure 1). Those overweight at 14 y (Figure 1) had a higher mean BMI and faster rate of increase than others, and those from a manual social class background had a higher mean BMI and a faster rate of increase than those from a non-manual background (Figure 2). Similar patterns were observed when considering the percentage over- Figure 1 Mean body mass index at four ages between 20 and 43 y for men and women, by weight at 14 years of age. Overweight de ned as relative weight of 120 (20% above the sex, age and height speci c standard weight) and above. 727 Results Sample size Every survey member with complete risk factor and confounding factor information, who had at least one recorded value of adult BMI, was included in the analyses (n ˆ 2659). The number with a recorded BMI at each point decreased with age (Table 1) and 1408 had all four values. Table 1 Number of observations at each pair of time points Age (y) * * * * Total number of individuals ˆ *Number of observations at each age (20, 26, 36 and 43 y). Figure 2 Mean body mass index at four ages between 20 and 43 y for men and women, by childhood social class. Childhood social class de ned in terms of father's occupation when survey members were aged 4 y.

4 728 weight ( > 25 kg=m 2 ) and obese ( > 30 kg=m 2 ) at each age (Table 2). There were a greater percentage of men than women overweight at each age, but a lower percentage obese. At all ages those from a manual social class had a greater proportion classi ed as overweight and obese compared with those from non-manual social classes. Of those overweight at 14 y, 40% were also overweight at 20 y and by 43 y 80% were overweight. The increase in obesity was also high in this group increasing from 6% at 20 y to 35% at 43 y. Statistical modelling Figures 1 and 2 provide an idea of the patterns of change in mean BMI and multilevel models provide estimates of the change after adjustment for all other relevant factors. Models provide estimates of the intercept, that is the mean BMI for different subgroups at age 20 y, and also of the slope, that is the rate of change of mean BMI for each yearly increase in age. In the simple age and sex model (Table 3, model (1)), for men at age 20 y, the estimated mean BMI (intercept) was kg=m 2, while for women it was kg=m 2 ; that is kg=m 2 minus 0.61 kg=m 2. Both linear (P < 0.001) and quadratic (P < 0.001) effects of age were signi cant. The patterns of change in BMI were, however, different for men and women, as indicated by signi cant sex by age interaction terms (Table 3, model (1)). The estimated mean linear increase in BMI was greater among men, at 0.12 kg=m 2 per year, than women for whom it was 0.03 kg=m 2 ; that is 0.12 kg=m 2 per year minus 0.09 kg=m 2 per year (Table 3, model (1)). However, the quadratic increase was greater among women than among men. These effects translate to women having a mean BMI of 0.61 kg=m 2 less than men at 20 y, approximately 1 kg=m 2 less at both 26 and 36 y with the difference decreasing again to 0.6 kg=m 2 less at 43 y. Birthweight was positively associated with adult BMI (0.53 kg=m 2 for every kg increase in birthweight), but its effect was cancelled out once relative weight at 7 y was added (not shown). All the other risk factors (weight at 7, 11 and 14 y and social class) were signi cantly associated with BMI (P < 0.05) and remained independently signi cant at the 5% level on addition of further factors (Table 3, models (2) ± (5)). There was a positive linear association between each childhood relative weight measure and BMI in adulthood, although the earlier measures were attenuated as later measures were added. For example, the effect of relative weight at age 7 y was reduced, on addition of relative weight at 11 and 14 y, from a 1.36 kg=m 2 (95% con dence interval (CI): 1.26 ± 1.46 kg=m 2 ) increase in BMI for every 10 units of relative weight (Table 3, model (2)) to a 0.14 kg=m 2 (95% CI: 0.02 ± 0.26 kg=m 2 ) increase (Table 3, model (4)). In the model including all relative weight measures (Table 3, model (4)), the measure at 14 y showed the strongest association, with an approximate 1kg=m 2 increase in BMI for every 10 units of relative weight. Survey members whose father's job was classi ed as manual had a mean BMI which was 0.53 kg=m 2 (95% CI: 0.37 ± 0.96 kg=m 2 ) higher than those whose father's job was classi ed as non-manual after taking account of childhood weight (Table 3, model (5)). The estimates of the effects of the relative weight measures were not confounded by social class. In terms of the random effects, the extent of random variation in the intercepts (s 2 u ) decreased as additional covariates were added to the model; from 5.83 in model (1) to 2.50 in model (5) (Table 3). Decreases in the variation were particularly seen as the childhood relative weights were added. In addition, the variance parameter estimates (s 2 u and s2 v in Table 3) were used to calculate expected ranges for the intercept and slope. For example, from the model including all signi cant childhood main effects (Table 3, model (5)), the 95% expected range for the intercept for men Table 2 Numbers and percentages overweight* and obese { at each adult age by categories of gender, father's social class and relative weight at 14 years Age (y) Number(%) Men Overweight 134 (11.5) 239 (21.9) 417 (42.2) 555 (57.9) Obese 8 (0.7) 21 (1.9) 60 (6.1) 113 (11.8) Women Overweight 113 (10.1) 182 (16.2) 236 (24.6) 379 (40.1) Obese 16 (1.4) 30 (2.7) 62 (6.5) 127 (13.4) Father, non-manual Overweight 78 (8.1) 124 (13.6) 207 (25.2) 336 (41.9) Obese 6 (0.6) 12 (1.3) 36 (4.4) 67 (8.4) Father, manual Overweight 169 (12.8) 297 (22.8) 446 (39.7) 598 (54.3) Obese 18 (1.4) 39 (3.0) 86 (7.7) 173 (15.7) Not overweight at 14 y Overweight 100 (5.2) 250 (13.5) 436 (26.8) 695 (43.4) Obese 3 (0.2) 14 (0.8) 47 (2.9) 134 (8.4) Overweight at 14 y { Overweight 147 (40.4) 171 (46.2) 217 (68.0) 239 (79.4) Obese 21 (5.8) 37 (10.0) 75 (23.5) 106 (35.2) *Overweight de ned as BMI > 25 kg=m 2. { Obese de ned as BMI > 30 kg=m 2. { Overweight at 14 y de ned as greater than or equal to 120 units of relative weight (20% above the sex, age and height speci c standard weight).

5 729 Table 3 Estimated mean differences (standard errors*) in BMI (kg=m 2 ) for given risk factors and the estimated components of between-individual and within-individual variation for 7 models Model (1) (2) (3) (4) (5) (6) (7) Fixed parameters Intercept { Sex (baseline ˆ men) women 0.61 (0.11)* 0.65 (0.10)* 0.55 (0.09)* 0.59 (0.09)* 0.60 (0.08)* 0.60 (0.08)* 0.62 (0.09)* Age per y 0.12 (0.01)* 0.12 (0.01)* 0.12 (0.01)* 0.12 (0.01)* 0.12 (0.01)* 0.10 (0.01)* 0.10 (0.01)* SexAge per y (women) 0.09 (0.02)* 0.09 (0.02)* 0.09 (0.02)* 0.09 (0.02)* 0.09 (0.02)* 0.09 (0.02)* 0.09 (0.02)* Age 2 per y (0.0004)* (0.0004)* (0.0004)* (0.0004)* (0.0004)* (0.0004)* (0.0004)* SexAge 2 per y 2 (women) (0.0006)* (0.0006)* (0.0006)* (0.0006)* (0.0006)* (0.0006)* (0.0006)* Relative weight at 7 y per 10 units Ð 1.36 (0.05)* 0.38 (0.06)* 0.14 (0.06) { 0.15 (0.06) { 0.13 (0.06) { 0.14 (0.06) { Relative weight at 11 y per 10 units Ð Ð 1.06 (0.04)* 0.42 (0.05)* 0.42 (0.05)* 0.42 (0.05)* 0.42 (0.05)* Relative weight at 14 y per 10 units Ð Ð Ð 0.99 (0.04)* 0.98 (0.04)* 0.80 (0.06)* 0.80 (0.06)* Father's social class (baseline ˆ non-manual) manual Ð Ð Ð Ð 0.53 (0.08)* 0.44 (0.08)* 0.25 (0.09)* Relative weight (14 y) age per y (per 10 units) 0.01 (0.002)* 0.01 (0.002)* Father's social classage per y (manual) Ð Ð Ð Ð Ð 0.03 (0.007)* 0.03 (0.007)* Sexrelative weight (14 y) per 10 units (women) Ð Ð Ð Ð Ð 0.26 (0.06)* 0.26 (0.06)* Education (baseline ˆ no quali cations) below `O' level Ð Ð Ð Ð Ð Ð 0.16 (0.15) `O' level 0.37 (0.11)* `A' level 0.46 (0.10)* degree 0.62 (0.15)* Level two random parameters s 2 u 5.83 (0.21) 4.33 (0.17) 3.28 (0.14) 2.55 (0.12) 2.50 (0.12) 2.49 (0.12) 2.46 (0.12) s 2 v 0.02 (0.001) 0.02 (0.001) 0.02 (0.001) 0.02 (0.001) 0.02 (0.001) 0.02 (0.001) 0.02 (0.001) suv 0.05 (0.009) 0.04 (0.008) 0.03 (0.007) 0.02 (0.007) 0.02 (0.007) 0.02 (0.007) 0.02 (0.007) Level one random parameters s 2 e 1.90 (0.05) 1.90 (0.05) 1.91 (0.05) 1.91 (0.05) 1.91 (0.05) 1.91 (0.05) 1.91 (0.05) * P < 0.01 for xed effects. { 0.01 < P < 0.05 for xed effects. { The intercept is the estimated mean BMI for the group with baseline values for every factor included in the model at age 20 y; that is men, relative weight of 100 (standard) at 7, 11 and 14 y, non-manual social class in childhood, no educational quali cations. s 2 u ˆ variance of the between-individual random variable associated with intercept u j. s 2 v ˆ variance of the between-individual random variable associated with slope v j. suv ˆ covariance of uj, vj. s 2 e ˆ variance of the within-individual random variable e ij.

6 730 (of non-manual social class and relative weight of 100 units at 7, 11p and 14 y) was (18.87 ± 25.07) kg=m 2 (21:97 1:96 2:50), while the corresponding range for women was (18.27 ± 24.47) kg=m 2. The interval for the slope for men (of non-manual social class and with a relative weight of 100 units at 14 y) was ( 0.16 ± 0.40) kg=m 2 per year. The positive covariance (s uv ) estimate between the slope and the intercept indicated that individuals with a high predicted BMI at age 20 y tended to show a larger increase in BMI over the period of interest than those with lower BMI at 20 y. An upper and a lower bound, around the average curve predicted by the xed part of the model, de ned a region within which, according to the model, approximately 95% of all individual curves lie. For men from a non-manual social class background, and with relative weight of 100 units at 7, 11 and 14 y, the plot indicated substantial variability and that this variability increased with age (Figure 3). The same increasing variability with age was observed for other subgroups as the model assumed that the betweenindividual variation at each age was the same for all groups. The within-individual variation was found to increase, as expected, with age and to vary by sex (results not shown). Signi cant interactions were observed between 14- year-old relative weight and age (P < 0.001), and father's social class and age (P < 0.001) (Table 3, model (6)). The higher the relative weight at 14 y, the greater the rate of increase in mean BMI per year (0.01 kg=m 2 greater for every 10 units of relative weight). Furthermore, those from a manual social class background had a rate of increase in BMI 0.03 kg=m 2 per year (95% CI: 0.02 ± 0.04 kg=m 2 ) greater than those from a non-manual social class. This translates to a widening of the difference between the two groups of 0.3 kg=m 2 every 10 y. The main effect of social class (0.44 kg=m 2 ) in this model now represents the difference between the two social class groups at age 20 y. The interaction between relative weight at 14 y and sex (Table 3, model (6)) indicated that the association between relative weight at 14 y and adult BMI was stronger in women than men. The effect of 10 units of relative weight on mean BMI at all ages was 0.26 kg=m 2 (95% CI: 0.14 ± 0.38 kg=m 2 ) greater among women than men. The effect of childhood social class was weakened on addition of the education variable (Table 3, model (7)), with the estimated mean difference between manual and non-manual social classes being reduced to 0.25 kg=m 2 (95% CI: 0.07 ± 0.43 kg=m 2 ), but was still highly signi cant (P < 0.01). Furthermore, the difference in change in BMI between the two classes remained unchanged (Table 3, models (6) and (7)). Figure 4 provides a plot of the adjusted estimated effects (using model (7)) of social class and being overweight at 14 y for men and women. Being overweight at 14 y had a greater in uence on adult BMI than social class, with the difference in mean BMI between overweight and average weight children being evident from the age of 20 y. The gap between the social classes (for each weight group) widened with increasing age (Figure 4). The greater curve in the change in BMI for women compared with men is evident, as is the stronger effect of relative weight among women (Figure 4). Educational quali cations were strongly associated with BMI, with those having higher levels of education having lower mean BMI (Table 3). Social class in adulthood did not exhibit a relationship with BMI in addition to the childhood risk factors (P ˆ 0.64) and neither was it associated with BMI after allowing for the stronger effect of education. The results from the analysis of the 1408 survey members with complete data were very similar to those from the main analysis and the overall conclusions did not change. The only substantive difference was the weakening of the effect of relative weight at 7 y after the addition of relative weight at 14 y with the effect becoming non-signi cant (P ˆ 0.7). In the nal model (model (7)), for 10 units increase in relative weight at 7 y the increase in BMI was only 0.08 kg=m 2 compared with 0.14 kg=m 2 in the main analysis. Discussion Figure 3 Average curve predicted by the xed part of model (5) for men (with baseline values for all risk factors) and the region within which approximately 95% of the individual curves for this subgroup would lie. Baseline values for risk factors are nonmanual social class in childhood, relative weight of 100 (standard) at ages 7, 11 and 14 y. The results indicated a continuing in uence of childhood social and biological factors on adult relative weight up until the age of 43 y in both men and women. From a methodological perspective, the results observed may have been dependent on how and when BMI was obtained. The self-reported

7 Figure 4 (a) Estimated differences in body mass index between subgroups relative to men at age 20 with relative weight of 100 (standard) at 14 y and from a non-manual social class in childhood. Overweight de ned as relative weight of 120 (20% above the sex, age and height speci c standard weight). (b) Estimated differences in body mass index between subgroups relative to women at age 20 with relative weight of 100 (standard) at 14 y and from a non-manual social class in childhood. Overweight de ned as relative weight of 120 (20% above the sex, age and height speci c standard weight). heights and weights at ages 20 and 26 y may have led to an underestimate of mean BMI. 15 The magnitude of the bias may, however, be small as it has been found that self-reported weight and height are remarkably accurate indicators of actual height and weight. 31 The implication is therefore that the increase in BMI observed with increasing age in these analyses may be a slight overestimate of the true rate of increase. If the extent of underreporting varied by subgroup, for example women have been found to underreport their weight to a greater degree than men, 31 then the estimated differences in change in BMI between these subgroups may have been biased. In terms of gender differences, however, the largest increase in BMI for women relative to men occurs between 36 and 43 y, ages at which height and weight were measured by trained nurses. It was, of course, not possible to model the uctuations in BMI in the periods between measurement occasions and it may be that the estimated increase in mean BMI in these analyses would have been modelled differently, for example using more complex curves, if measurements had been repeated more often. High relative weight in childhood was associated with high BMI in adulthood. The effect in relation to relative weight at age 14 y was stronger among women than men and also became stronger with increasing age for both sexes. It could be hypothesised that a social pathway exists between childhood relative weight and adult BMI, as it has been shown that severely obese young men are less likely to attain a high social class and high educational quali cations 37,38 and obese young women are less likely to achieve high earnings. 39 However, these ndings of social disadvantage relate only to the obese, rather than showing increasing disadvantage with increasing BMI. Furthermore, the effects of the childhood relative weight were independent of educational attainment and adult social class in this study. Alternatively, high childhood relative weight may be an indicator of a genetic predisposition towards higher BMI and subsequently to obesity throughout life. This may be indicated by the independent effects of relative weight at 7, 11 and 14 y, where it is those with high relative weight at every age who had the highest mean BMI in adulthood. The Bogalusa Heart study 40 showed that the offspring of parents with early cardiovascular disease were consistently overweight, beginning in childhood, and they conclude that both genetic predisposition to later heart disease and environment are important factors in uencing this nding. As in other studies, 24,25 and at an earlier stage in this study, 17 an association between childhood social circumstances and adult BMI which is independent of adult social class was observed. In addition, the increase in BMI was greater in those from the manual social classes than in those from the nonmanual social classes of origin. Childhood circumstances may in uence adult obesity through the longterm effect of lifestyle habits such as diet and exercise acquired in childhood 17 Adults of lower socioeconomic status and those with lower educational quali- cations, in this study and others, have less healthy diets 41,42 and are less likely to participate in sports activities. 43,44 Hence, a gradual accumulation of weight from early adulthood as a result of poor health-related behaviours throughout the life course may lead to the social inequalities, in terms of childhood social class, in BMI. Furthermore, individuals born into a non-manual social class are more likely to achieve a high socioeconomic status and high earnings as adults 45 ± 49 and higher levels of educational quali cations. 28,46,50 The inability to move out of the lower social classes may continue to limit opportunities for changing lifestyle behaviours. 731

8 732 Although obesity has been found to be associated with childhood social class, behavioural factors such as diet and exercise have been found to be related to adult social class or education. 24,25,41,43 Additionally, those of low socioeconomic status in adulthood are more likely to smoke, 51 but smokers have been found to have lower BMI than non-smokers. 52 Adjustment for education and social class in the current analysis may account, at least to some extent, for the social pathway linked to health-related behaviours. Education was indeed found to account for some, but not all, of the childhood social class effect. Further analyses (not shown), where the nal model in our study was adjusted for measures of adult smoking and exercise habits, indicated that those who were smokers and those who exercised regularly at both 36 and 43 y had lower mean BMI across all ages compared with others. These adjustments strengthened the association between education and BMI, but did not confound the effect of childhood social class. An alternative explanation for the remaining association between childhood social class and BMI observed in the present study was therefore a possibility. Results from the study of the Dutch famine of 1944 ± suggest that nutritional deprivation in the rst half of pregnancy resulted in higher obesity rates at age 19 y. Hence, social class in the current study may have provided an indication of nutritional status during pregnancy. However, the post-war food rationing at that time may have meant that there was little variation in intake across the social classes. Although this may have been true for many foods, one-day records of meals eaten by the survey members at the age of 4 y indicated that there was a social class variation in relation to particular foods, including vegetables and fruits. 54 There was no independent effect of birthweight, which was our only direct marker of the fetal environment, on BMI in our study. However, programming effects in terms of central adiposity have been suggested by other studies, with low birthweight being correlated with high adult waist ± hip ratio, 21,22 even though birthweight is positively correlated with BMI. The NSHD, however, only has one measure of waist ± hip ratio (at 43 y), and hence the current analysis could not be repeated with this measure of central adiposity as the outcome. The estimates of the random variation from the multilevel models indicate a wide range of patterns of change across this sample, indicating that there is substantial variation around the mean pattern of increasing BMI with increasing age. The quanti cation of the components of variation, and their change across different models, are interesting in themselves as they provide additional information regarding the data. Risk factors other than those considered in the main analysis will explain some, but are unlikely to explain all, of the remaining variation between individuals. For example, smoking and exercise each accounted for a small amount. Among women, parity was an obvious additional factor to consider, and analyses (not shown) indicated that it explained a further small amount of the random variation. The analyses indicated that women with 3 or more children had higher BMI than women with less or no children, while parity did not attenuate the association between childhood social class and adult BMI. A speci c feature of the data used here was that the individuals included in the analysis were all of the same age. Hence, the effect of age was longitudinal and could only be generalised to this cohort and a speci c cohort effect could not be ruled out. For example, the relationships observed could have been in uenced by the experience of post-war food rationing in the early years of life. The use of multilevel models to analyse these data means that the relationships between risk factors and outcome at different ages can be tested simultaneously, while appropriately taking into account the correlation between the measures at different occasions. This uni ed approach is preferable to considering the effects of the childhood factors on each outcome measure separately, where results cannot be compared formally across models. In order for the estimates of the parameters of interest to be unbiased, it must be assumed that responses are missing at random as there is evidence that the models used here may perform badly if this assumption does not hold. 55 In our study, for example, if it were the case that obese cohort members were more likely to have missing BMI measures, because they did not want to be weighed or because they were too ill to participate in data collection, then missing at random (MAR) could not be assumed. However, a test of `missing completely at random' (MCAR) 56 was carried out and provided no evidence against MCAR, which is a stronger assumption than MAR. Furthermore, a model predicting outcome missingness suggested that it was not dependent on the level of the previous BMI measure or on any of the childhood risk factors. Hence, it was thought reasonable to assume MAR in these data. In conclusion, childhood relative weight and childhood social class were shown to have an effect on adulthood BMI and change in BMI from ages 20 to 43 y, even after adjustment for adult socioeconomic status and education. Some continuity between childhood and adult relative weight up to the age of 43 y was observed in this cohort who experienced post-war food rationing in their early years. This is of public health concern to later cohorts, where childhood relative weight and the prevalence of childhood obesity are higher. The in uence of childhood social class, which was only partially confounded by adult socioeconomic status, on adult BMI and change in BMI suggests that both behavioural and biological processes may be operating from early life. The social and biologic pathways underlying these observations are not yet fully understood.

9 Acknowledgements This research was funded by the UK Medical Research Council. References 1 Hubert HB, Feinleib M, McNamara PM, Castelli WP. Obesity as an independent risk factor for cardiovascular disease: a 26 year follow-up of participants in the Framingham heart study. Circulation 1983; 67: 968 ± Gar nkle L. Overweight and cancer. Ann Intern Med 1985; 103: 1034 ± Le Marchand L, Kolonel LN, Earle ME, Mi M-P. Body size at different periods of life and breast cancer risk. Am J Epidemiol 1988; 128: 137 ± Manson JE, Colditz GA, Stampfer MJ, Willett WC, Rosner B, Monson RR, Speizer FE, Hennekens CH. A prospective study of obesity and risk of coronary heart disease in women. N Engl J Med 1990; 322: 882 ± Colditz GA, Willett WC, Stampfer MJ, Manson JE, Hennekens CH, Arky RA, Speizer FE. Weight as a risk factor for clinical diabetes in women. Am J Epidemiol 1990; 132: 501 ± Walker M, Wannamethee G, Whincup PH, Shaper AG. Weight change and risk of heart attack in middle-aged British men. Int J Epidemiol 1995; 24: 694 ± Rimm EB, Stampfer MJ, Giovannucci E, Ascherio A, Spiegelman D, Colditz GA, Willett WC. Body size and fat distribution as predictors of coronary heart disease among middle-aged and older US men. Am J Epidemiol 1995; 141: 1117 ± Willett WC, Manson JE, Stampfer MJ, Colditz GA, Rosner B, Speizer FE, Hennekens CH. Weight, weight change, and coronary heart disease in women: risk within the `normal' weight range. JAMA 1995; 273: 461 ± Lew EA, Gar nkle L. Variations in mortality by weight among 750,000 men and women. J Chronic Dis 1978; 32: 563 ± Wannamethee G, Shaper AG. Body weight and mortality in middle aged British men: impact of smoking. BMJ 1989; 299: 1497 ± Lee I-M, Manson JE, Hennekens CH, Paffenbarger RS. Body weight and mortality. A 27 year follow-up of middle-aged men. JAMA 1993; 270: 2823 ± Kushner RF. Body weight and mortality. Nutr Rev 1993; 51: 127 ± Manson JE, Willett WC, Stampfer MJ, Colditz GA, Hunter DJ, Hankinson SE, Hennekens CH, Speizer FE. Body weight and mortality among women. N Engl J Med 1995; 333: 677 ± Gunnell DJ, Frankel SJ, Nanchahal K, Peters TJ, Davey Smith G. Childhood obesity and adult cardiovascular mortality: a 57- y follow-up study based on the Body Orr cohort. Am J Clin Nutr 1998; 67: 1111 ± Stark O, Atkins E, Wolff OH, Douglas JWB. A longitudinal study of obesity in the National Survey of Health and Development. BMJ 1981; 283: 13 ± Whitaker RC, Wright JA, Peper MS, Seidel KD, Dietz WH. Predicting obesity in young adulthood from childhood and parental obesity. N Engl J Med 1997; 337: 869 ± Braddon FM, Rodgers B, Wadsworth MEJ, Davies JMC. Onset of obesity in a 36 year birth cohort study. BMJ 1986; 293: 299 ± Barker DJP, Ed. Fetal and Infant Origins of Adult Disease. BMJ Publishing Group: London, Barker DJP. Mothers, Babies and Disease in Later Life. BMJ Publishing Group: London, Leon DA, Koupilova I, Lithell HO, Berglund L, Mohsen R, Vagero D, Lithell U-B, McKeigue PM. Failure to realise growth potential in utero and adult obesity in relation to blood pressure in 50 year-old Swedish men. BMJ 1996; 312: 401 ± Fall CHD, Osmond C, Barker DJP, Clark PMS, Hales CN, Stirling Y, Meade TW. Fetal and infant growth and cardiovascular risk factors in women. BMJ 1995; 310: 428 ± Law CM, Barker DJP, Osmond C, Fall CHD, Simmonds SJ. Early growth and abdominal fatness in early life. J Epidemiol Community Health 1992; 46: 184 ± Power C, Moynihan C. Social class and changes in weight-forheight between childhood and early adulthood. Int J Obes 1988; 12: 445 ± Blane D, Hart CI, Davey Smith G, Gillis CR, Hole DJ, Hawthorne VM. Association of cardiovascular risk factors with socioeconomic position during childhood and during adulthood. BMJ 1996; 313: 1434 ± Wannamethee SG, Whincup PH, Shaper G, Walker M. In uence of fathers' social class on cardiovascular disease in middle-aged men. Lancet 1996; 348: 1259 ± Goldstein H. Multilevel Statistical Models (2nd edition). Edward Arnold: London, Diggle PJ, Liang K-L, Zeger SL. Analysis of Longitudinal Data. Oxford University Press: Oxford, Wadsworth MEJ. The Imprint of Time: Childhood, History and Adult Life. Oxford University Press: Oxford, Wadsworth MEJ, Kuh DJL. Childhood in uences on adult health: a review of recent work in the British 1946 national cohort study, the MRC National Survey of Health and Development. Paediatr Perinat Epidemiol 1997; 11: 2 ± Wadsworth MEJ, Mann SL, Rodgers B, Kuh DJL, Hilder WS, Yusuf EJ. Loss and representativeness in a 43 year follow-up of a national birth cohort. J Epidemiol Community Health 1992; 46: 300 ± Stewart AL. The reliability and validity of self-reported weight and height. J Chronic Dis 1982; 35: 295 ± Newens EM, Goldstein H. Height, weight, and the assessment of obesity in children. Br J Prev Soc Med 1972; 26: 33 ± Cole TJ. A method for assessing age-standardized weight-forheight in children seen cross-sectionally. Ann Hum Biol 1979; 6: 249 ± Rasbash J, Woodhouse G. Mln Command Reference, version la. Multilevel Models Project, Institute of Education, University of London: London, Woodhouse G. Multilevel Modelling Applications: A Guide for Users of Mln. Multilevel Models Project, Institute of Education, University of London: London, Little RJA, Rubin DB. Statistical Analysis with Missing Data. Wiley: New York, Sonne-Holme S, Sorensen TIA. Prospective study of attainment of social class of severely obese subjects in relation to parental social class, intelligence, and education. BMJ 1986; 292: 586 ± Sorensen TIA. Socio-economic aspects of obesity: causes or effects? Int J Obes 1995; 19(Suppl 6): S6 ± S8. 39 Sargent ID, Blanch ower DG. Obesity and stature in adolescence and earnings in young adulthood. Arch Pediatr Adolesc Med 1994; 148: 681 ± Bao W, Srinivasan SR, Valdez R, Greenlund KJ, Wattigney WA, Berenson GS. Longitudinal changes in cardiovascular risk from childhood to young adulthood in offspring of parents with coronary artery disease. JAMA 1997; 278: 1749 ± Braddon FEM, Wadsworth MEJ, Davies JMC, Cripps HA. Social and regional differences in food and alcohol consumption and their measurement in a national birth cohort. J Epidemiol Community Health 1988; 42: 341 ± Blaxter M. Health and Lifestyles. Routledge: London, Kuh DJL, Cooper C. Physical activity at 36 years: patterns and predictors in a longitudinal study. J Epidemiol Community Health 1992; 46: 114 ± Of ce of Population Censuses and Surveys. General Household Survey Series GHS, No. 16. HMSO: London Goldthorpe JH. Social Mobility and Class Structure in Modern Britain. Clarendon Press: Oxford,

10 Halsey AH, Health AF, Ridge JM. Origins and Destinations: Family, Class and Education in Modern Britain. Clarendon Press: Oxford, Britten N. Models of intergenerational class mobility: ndings from the National Survey of Health and Development. Br J Social 1981; 32: 224 ± Kuh D, Wadsworth M. Childhood in uences on adult male earnings in a longitudinal study. Br J Sociol 1991; 42: 537 ± Kuh D, Head J, Hardy R, Wadsworth M. The in uence of education and family background on women's earnings in midlife: evidence from a British national cohort study. Br J Social Education 1997; 18: 385 ± Douglas JWB. The Home and the School. MacGibbon & Kee: London, Wald N, Kiryluk S, Barby S, Doll R, Pike M, Peto R. U.K. Smoking Statistics. Oxford University Press: Oxford, Albanes D, Jones Y, Micozzi MS, Mattson ME. Associations between smoking and body weight in the US population analysis of NHANES II. Am J Public Health 1987; 77: 439 ± Ravelli G-P, Stein ZA, Susser MW. Obesity in young men after famine exposure in utero and early infancy. New Eng J Med 1976; 295: 349 ± Prynne CJ, Paul AA, Price GM, Wadsworth MEJ. Social class differences in foods eaten by a national sample of four-year old children in 1950 (abstract). Proceedings of the Nutrition Society (in press). 55 Wang-Clow F, Lange M, Laird NM, Ware JH. A simulation study of estimators for rates of change in longitudinal studies with attrition. Stat Med 1995; 14: 283 ± Park T, Lee S-Y. A test of missing completely at random for longitudinal data with missing observations. Stat Med 1997; 16: 1859 ± 1871.