点击显示 收起
1 From the Division of Epidemiology, School of Public Health, University of Minnesota, Minneapolis (KS); the Department of Public Health, University of Helsinki, Finland (KS, SS-L, EL, and JK); the Department of Public Health, University of Turku, Finland (MK); and the Department of Mental Health and Alcohol Research, National Public Health Institute, Finland (JK).
2 Supported by the Academy of Finland, Research Council for Health (grants 53585 and 52277), and the European Commission under the program "Quality of Life and Management of the Living Resources" of the 5th Framework Program (grant QLG2-CT-2002-01254). 3 Address reprint requests to K Silventoinen, Department of Public Health, University of Helsinki, PO Box 41, Mannerheimintie 172, FIN-00014, University of Helsinki, Helsinki, Finland. E-mail: karri.silventoinen{at}helsinki.fi.
ABSTRACT
Background: Disparities in body mass index (BMI) between persons with different educational levels in Western countries are well documented, but the background of these education-associated disparities remains poorly understood.
Objective: The objective was to examine the influence of environmental and genetic factors on education-associated disparities in self-reported BMI and weight change.
Design: Longitudinal postal surveys were performed in 1975, 1981, and 1990. The data were analyzed by using multivariate genetic models for twin data. The data derived from the Finnish Twin Cohort included 2482 monozygotic and 5113 dizygotic same-sex male and female twin pairs born between 1915 and 1957.
Results: Education-associated differences in BMI and in weight change were clear in 1975 and 1981, respectively, whereas no differences were seen in weight change between 1981 and 1990 when age and baseline BMI were adjusted for. The trait correlation between baseline BMI and educational attainment (–0.15 in men and women) was mainly due to correlations between additive genetic factors that contributed to BMI and education in men (–0.20; 95% CI: –0.25, –0.14) and women (–0.32; 95% CI: –0.40, –0.25) when adjusted for age. Among women, a weaker positive correlation was found for the unshared environmental effects contributing to the 2 traits (0.06; 95% CI: 0.02, 0.12). The same factors that affected the association between education and BMI in 1975 largely explained the association between education and weight change in 1981.
Conclusion: The results suggest the possibility that common genetic factors affect educational attainment and body weight, which contribute to education-associated disparities in BMI in adulthood.
Key Words: BMI • weight change • education • heritability • twins
INTRODUCTION
An inverse association between obesity measured by body mass index (BMI; in kg/m2) and various indicators of socioeconomic status such as educational attainment, social class, or income level is a common feature of affluent societies, especially among women (1, 2). Moreover, these differences tend to widen with age because socioeconomic status has been found to be inversely associated with weight gain (3–5). The background of socioeconomic variation in obesity is still poorly understood. Better comprehension of the causal mechanisms underlying socioeconomic disparities in obesity and weight gain would help to identify risk groups and formulate measures that might be used to reduce obesity among disadvantaged persons.
Three potential pathways underlying the socioeconomic variation of BMI can be distinguished. First, as shown by previous studies, social background may contribute because the socioeconomic status of origin is associated with adult BMI (6–8). Many factors in childhood, such as parental neglect (9), poor area of residence (6), and social deprivation (10) have been found to be associated with obesity and are probably more prevalent in lower socioeconomic positions. Second, a direct association between BMI and socioeconomic status is possible, because both men and women of higher social classes appear to be more concerned about obesity, healthy eating, and physical exercise (11). Moreover, low socioeconomic status may increase stress, which appears to be positively associated with weight gain (12). Third, common genetic factors may affect both BMI and socioeconomic status. Previous studies have shown that BMI is highly heritable (13, 14), and there is also a genetic component underlying socioeconomic differences as measured by educational attainment (15, 16). It is possible that some of the genetic factors predisposing to BMI may also account for genetic variation in educational attainment.
A twin study design is a powerful tool for distinguishing the genetic and environmental factors influencing a phenotype or trait variation. In this study we examined the effect of genetic and environmental factors on disparities in BMI among a large adult Finnish twin data set with different educational levels. We also investigated whether genetic and environmental factors contributed to education-based differences in weight change between 1975 and 1981 and between 1981 and 1990. Our data allowed us to compare the background of education-based differences in BMI and weight gain between men and women. To our knowledge, this kind of approach has not been previously applied to examining variation in BMI and weight gain between people with different levels of education.
SUBJECTS AND METHODS
The data were derived from the Finnish Twin Cohort Study (12, 17). The first questionnaire was mailed in 1975 to all same-sex Finnish twin pairs born before 1958 with both cotwins alive in 1974. The response rate to this questionnaire was 89%. Follow-up questionnaires were sent in 1981 and 1990, with response rates of 84% and 77%, respectively. In the 1990 survey, the questionnaire was not sent to twins born before 1930, and both cotwins also needed to be alive. All complete twin pairs who replied to the questionnaires in 1975 and 1981 with information about education and BMI from both twins were included in this study.
Twin zygosity was probed in the 1975 survey with questions about the similarity of appearance of a twin pair at an early school age. The reliability of this questionnaire method for classifying zygosity was assessed by using 11 blood markers in a subsample of 104 twin pairs classified as monozygous or dizygous (18). The agreement between the results of the blood test and the questionnaire was 100%. The probability of misclassification of a twin pair in this subsample was estimated at 1.7%.
Body height and weight were questioned for identically in each questionnaire (19). Weight was given in kilograms and height in centimeters; when necessary, the values were rounded to the nearest integral number. The validity of self-reported height and weight was controlled in a subsample of twins who had replied to the questionnaire in 1990 (n = 100 men and 127 women). A clinical examination after the questionnaire showed that the correlation between self-reported and measured BMI was 0.89 among men and 0.90 among women, which suggested good reliability of self-reported BMI (12). Weight change was computed between weights reported in 1975 and 1981 and between 1981 and 1990.
The 504 twin pairs who could not be classified by zygosity were excluded from the data. The final numbers of respondents were 2482 monozygous and 5113 dizygous twin pairs in the analysis concerning BMI in 1975 and concerning weight changes by 1981 and in 1317 monozygous and 2417 dizygous twin pairs in the analysis concerning weight changes by 1990.
Data on education were self-reported. Because the youngest respondents had not yet completed their final education in 1975, we based educational attainment on the data from the 1981 questionnaire. The respondents were asked to classify themselves in 1 of 8 categories by the length of their education: less than primary school (3 y of education), primary school (6 y of education), junior high school (9 y), high school graduate (12 y), university degree (16 y), and 1 y of education such as vocational training in addition to primary school (7 y total), junior high school (10 y), or high school (13 y). In the genetic modeling, years of education were used. The kappa () coefficient for education between the 1981 and 1975 surveys was 0.74 for both men and women when respondents younger than 30 y in 1975 were omitted, which indicated good reliability of the self-reported education data (15).
Twin analysis based on linear structural equations was used (20). A classic twin design is based on the assumption that dizygous twins share on average 50% of their segregating genes, whereas monozygous twins are genetically identical. Environmental variation is divided into variation due to shared and unshared environment, with correlations of 1 and 0, respectively, within both monozygous and dizygous twin pairs. Thus, by definition, the shared environment includes all environmental factors that make the twin pair similar for the trait and the unshared environment includes all environmental factors that make it dissimilar. On the basis of these assumptions, it is possible to estimate additive genetic variance (A) and nonadditive genetic variance (D), which consists of the interaction of allele effects at the same locus (dominance) as well as shared environmental (C) and unshared environmental (E) components of variance. In genetic modeling these variance components are treated as latent and standardized independent variables, which are used to explain the variation of the trait treated as the dependent variable in the model. Thus, the regression coefficients of these latent variables are the square roots of the genetic and environmental variance components A, D, C, and E affecting the trait. In this study the results are presented as the proportion of the trait variation explained by the genetic and environmental variation. The twin model further assumes random mating with respect to the traits in question and the absence of gene-gene and gene-environment interactions. Because our data included only twin pairs reared together, but not adopted twins or other relatives, we were not able to estimate simultaneously the effects of C and D.
In the genetic modeling, we first studied which model was needed to describe the trait variation of education and BMI. This model was used in bivariate analyses for education and BMI in 1975 (Figure 1). Bivariate modeling allowed us to estimate correlations between environmental and genetic factors affecting education and BMI and thus to determine whether the trait correlation between education and BMI is partly or completely due to the same or linked genetic factors or to the same or correlated environmental factors shared or unshared by twins. The statistical significance of these correlations was studied by fitting nested models and examining the change in the chi-square values (2) between the models with and without the correlation parameter (paths g12 for genetic factors and e12 for specific environmental factors in Figure 1) set to zero.
FIGURE 1.. Simplified version of bivariate and trivariate models including genetic (G1, G2, and G3) and specific environmental variance components (E1, E2, and E3) that affected education (Edu) and body mass index in 1975 (BMI75) and 1981 (BMI81) through corresponding path coefficients (g12, g13, and g23 for genetic and e12, e13, and e23 for specific environmental factors).
The analyses were continued by examining whether the genetic and environmental correlations behind the association between weight change and education differed from the correlations behind the association between education and BMI at baseline. We first analyzed the weight change between 1975 and 1981 using trivariate modeling with BMI in 1975, education, and BMI in 1981 and testing the statistical significance of the respective correlations between education and BMI in 1981. This model tested whether the genetic and environmental correlations between education and BMI in 1981 (paths g13 for genetic factors and e13 for specific environmental factors in Figure 1) were statistically significant, allowing for the correlations between education and BMI in 1975 (paths g12 and e12) as well as BMI in 1975 and 1981 (paths g23 and e23). Correspondingly, we analyzed weight change between 1981 and 1990 using a trivariate model that included education and BMI in 1981 and 1990. The modeling used the Cholesky decomposition parameterization (20) and was carried out with the Mx statistical package, version 1.54, designed for quantitative studies of twin and family data (21). Chi-square and root mean square error of approximation (RMSEA) statistics were used to assess the model fits (22).
In the descriptive analysis, the means of BMI and weight change by educational categories were adjusted for age by variance analysis with the use of the SAS for WINDOWS statistical package, version 8.02 (23). The statistical significance of the age-adjusted education trend of BMI and weight change was computed by using the same method. We also tested whether these associations were independent of age by analyzing the associations between BMI and education among the 18–29-y-olds (birth cohorts 1947–1957) and 30–54-y-olds (birth cohorts 1915–1946) at baseline in 1975. Because both BMI and education were highly correlated with age, age-adjusted residuals of BMI and education were used when calculating correlation matrices and the covariance matrices used in the genetic modeling. The residuals were calculated separately for men and women. A logarithmic transformation of BMI was used throughout to normalize the distribution.
RESULTS
Mean BMIs in 1975 and weight changes between 1975 and 1981 and between 1981 and 1990 are shown in Table 1. The basic results for BMI and weight change were reported earlier (24, 25). Education-based variation in the mean BMI in 1975 was clear and statistically significant among men (P < 0.001) and women (P < 0.001). The age-adjusted BMI in 1975 was highest among men with a primary school education (24.0) and among women with less than a primary school education (23.2) and was lowest among men and women with a university degree (22.7 and 21.3, respectively). Weight gain by 1981 was greatest among men and women with less than a primary school education (3.13 and 2.76 kg, respectively) and least among men with a university degree (1.92 kg) and women with more than a high school education (1.34 kg) when adjusted for age. The education-based trend in weight change was statistically significant among men (P = 0.054) and women (P = 0.002). As expected, the SDs for weight change were large. By 1990, weight had increased in all education groups compared with 1981 values; the education-based pattern in weight change between 1981 and 1990 was statistically significant among women (P = 0.021) but not among men (P = 0.522).
View this table:
TABLE 1. BMI in 1975 and weight changes between 1995 and 1981 and between 1981 and 1990 by sex, zygosity, age, and education level1
Age-adjusted partial Pearson correlation coefficients of education and BMI within monozygous and dizygous twin pairs are shown in Table 2. All correlations were statistically highly significant (P < 0.001). The trait correlations for education and for BMI in 1975, 1981, and 1990 ranged from 0.65 to 0.85 for monozygous and from 0.30 to 0.62 for dizygous twins. In the cross-trait correlations, education was negatively and BMI was positively correlated with the cotwin’s BMI.
View this table:
TABLE 2. Correlations of education and BMI in 1975, 1981, and 1990 adjusted for age within monozygous and dizygous twin pairs (twin A and twin B) by sex1
Educational differences in BMI and weight change were tested by computing partial Pearson correlation coefficients adjusted for age. The correlation between BMI in 1975 and education was –0.15 for both men and women. Partial correlations between education and weight change by 1981 were –0.04 for men and –0.06 for women when age and BMI in 1975 were adjusted for. All of these correlations were statistically highly significant (P < 0.001), although relatively modest in size. The correlations between education and weight change from 1981 to 1990 were statistically nonsignificant among men (0.015; P = 0.350) and women (–0.006; P = 0.687) when age and BMI in 1981 were adjusted for.
Genetic modeling began by estimating the best model for education and BMI with the use of the univariate model (data not shown). When modeling BMI, the additive genes/unique environment (AE) model gave the best fit in men (2 = 3.9, P = 0.42 in 1975; 2 = 1.6, P = 0.82 in 1981; and 2 = 3.5, P = 0.47 in 1990; df = 4 in all models). Fitting a more complex model, ie, the additive genes/dominant genes/unique environment (ADE) or the additive genes/common environment/unique environment (ACE) model, did not improve the fit of the model statistically significantly. In women, the ADE model had the best fit for BMI in 1975 (2 = 2.2, P = 0.53, df = 3) and 1981 (2 = 7.6, P = 0.06, df = 3), but the AE model had a better fit in 1990 (2 = 1.5, P = 0.82, df = 4). For education, the ACE model was the only one that gave an adequate fit in both men (2 = 7.6, P = 0.06, df = 3) and women (2 = 1.8, P = 0.61, df = 3). In the subsequent bi- and trivariate modeling, we used the AE model for BMI in men, the ADE model for BMI in women, and the ACE model for education in both men and women as the starting point in the analyses.
The results for bi- and trivariate modeling for men are shown in Table 3, and the corresponding analyses for women are shown in Table 4. In the bivariate models between education and BMI in 1975, we found that the fit of the base model was poorer (men: 2 = 29.6, P = 0.005, df = 13; women: 2 = 32.1, P = 0.001, df = 12) than expected on the basis of the results of univariate modeling. However, this was partly due to the large sample size, which easily leads to high chi-square values. The RMSEA criteria, which are less sensitive to sample size than are chi-square statistics, for the base model were satisfactory in men (0.030) and women (0.026). Setting the additive genetic correlation to zero made the fit of the model statistically significantly poorer in men (2 = 53.7, P = 0.001, df = 1) and women (2 = 126.5, P = 0.001, df = 1), which indicated the importance of this parameter in the model. Among women, the unshared environmental correlation was also statistically significant (2 = 6.9, P = 0.008, df = 1).
View this table:
TABLE 3. Fit of model for bivariate and trivariate analyses of education, BMI in 1975, and the change () in BMI from 1975 to 1981 and from 1981 to 1990 among men1
View this table:
TABLE 4. Fit of model for bivariate and trivariate analyses of education, BMI in 1975, and the change in in BMI from 1975 to 1981 and from 1981 to 1990 among women1
In trivariate modeling, the RMSEA criteria were satisfactory in the models for education and BMI in 1975 and 1981 (0.023 in men and 0.022 in women) as well as for education and BMI in 1981 and 1990 (0.023 and 0.021, respectively). In the model for education and BMI in 1975 and 1981, of the correlations between variance components, only the unshared environmental correlation among men was statistically significant (2 = 4.0, P = 0.044, df = 1). An examination of the correlations in the trivariate model including education and BMI in 1981 and 1990 showed that the only statistically significant correlation was additive genetic correlation among men (2 = 4.2, P = 0.040, df = 1).
The modeling was continued by examining more closely the parameters in the bivariate models (Table 5). It was found that the additive genetic factors were the main determinants of the trait variation both for BMI in 1975 [heritability (h2) = 0.71 in men and 0.69 in women includes additive (0.46) and dominance (0.23) genetic variation] and for education (h2 = 0.45 and 0.46, respectively). The shared environment accounted for a substantial part of the variation in education. The effect of unshared environment was somewhat weaker than was the effect of genetic factors on both BMI and education among men and women. The negative correlation between the additive genetic components affecting BMI and education was strong among men (–0.20; 95% CI: –0.25, –0.14) and women (–0.32; 95% CI: –0.40, –0.25). A weaker, but still statistically significant, positive correlation between the unshared environmental factors (0.06; 95% CI: 0.02, 0.12) was found among women.
View this table:
TABLE 5. Proportion of trait variance explained by genetic and environmental factors in the bivariate model for BMI in 1975 and education, by year of birth among men and women1
Cohort differences in the model were examined by testing the significance of the change in the fit of the model when the parameters of the cohorts born in 1915–1946 and 1947–1957 were fixed to be the same. In practice, we compared the chi-square value of the model in which the parameter estimates of both birth cohorts were fixed as identical to the sum of the chi-square values of the cohort-specific models. We found that the requirement that the parameters be the same made the fit of the model statistically significantly poorer when the differences between the cohorts in men (2 = 58.9, P = 0.001, df = 7) and women (2 = 155, P = 0.001, df = 8) were examined, which suggested that the overall differences were not due to chance. However, CIs of the single parameters were wide and their differences by birth cohort were statistically nonsignificant. We did not conduct a similar formal heterogeneity test for sex differences because different models were used for BMI in men and women. However, in a comparison of single parameter estimates, only minor nonsignificant differences in their magnitude between men and women were observed.
Because the fit of the bivariate model was poorer than expected on the basis of the univariate modeling, we decided to examine in more detail which factors might account for the decrease in the fit (data not shown). When using the ACE model both for education and BMI in 1975, we found that the fit was substantially better (men: 2 = 18.7, P = 0.064, df = 11; women: 2 = 18.2, P = 0.077, df = 11) than when the AE or the ADE models were used (Tables 3 and 4, respectively). This improved fit was due to a statistically significant shared environmental correlation between education and BMI (men: 2 = 10.8, P = 0.001, df = 1; women: 2 = 25.5, P = 0.001, df = 1), despite the lack of shared environmental influences on BMI in the univariate models. A closer study of this unexpected finding showed that it was because the dizygous correlation between education and BMI in 1975 was greater than half of the monozygous correlation expected if only the additive genetic component affected this correlation (Table 2).
DISCUSSION
We found clear education-based differences in BMI among Finnish twins, which agrees with the findings of previous Finnish studies (26, 27). Education was associated with weight change between 1975 and 1981 but not between 1981 and 1990 when age and baseline BMI (1975 or 1981) were adjusted for. The data sets were not fully comparable between 1975 and 1981 and between 1975 and 1990, because in 1990 the questionnaire was not sent to the oldest twins. However, when the correlations between education and weight change among those who had taken part in all 3 surveys were studied, the correlations were still lower for the 1981 to 1990 comparison than for the 1975 to 1981 comparison. The lack of correlation between education and weight change between 1981 and 1990 may be associated with the aging of the participants or a period effect during these years. However, because education in the data was strongly correlated with age, and the association between education and weight change may well vary between birth cohorts in an unknown way, this question needs to remain open.
A genetic component behind both BMI and education was found, whereas the unshared environment had a smaller effect. The shared environment had an effect on education but not on BMI. These results agree with previous studies concerning the genetic architecture of BMI (13, 14) and education (16). A study of factors affecting the educational variation in BMI and weight changes between 1975 and 1981 showed that the same factors affecting the correlation between education and BMI at baseline were also likely to influence the correlation between education and weight change, with the possible exception of unshared environment among men. Previous studies have shown socioeconomic disparities in weight gain (3–5). Our results suggest that these differences, as measured by education, are largely associated with the same genetic and shared environmental factors that contribute to earlier educational differences in BMI.
On closer examination of the factors affecting the trait correlation between BMI and education, we found that the common genetic factors affecting both BMI and education are likely to be the main explanation for this correlation among both men and women. The role of genetic factors agrees with previous Danish results, where the obesity of adoptees in adulthood was found to be associated with the socioeconomic status of the biological fathers (28). Intelligence is one potential explanation for this association because it appears to be associated with BMI (29). Previous twin studies have shown that the heritability of intelligence is high, ie, between 0.60 and 0.80, whereas the effect of shared environment is minimal in adulthood and marked in childhood (30, 31).
We also found among women a positive unshared environmental correlation affecting the correlation between BMI and education. This indicates the likelihood of environmental factors not shared by the twin pair that would positively influence both BMI and education and thus attenuate rather than increase the education-based gradient in BMI. This is a surprising finding and because we found that among men this correlation was negative, even if statistically nonsignificant, it should be treated with caution. Nevertheless, the finding suggests that unshared environmental factors cannot be the main determinants underlying education-based differences in BMI. Thus, for example, better health consciousness among well-educated people (11) or stigmatization of obese children in educational institutions (32), both of which would tend to be included mainly in the unshared environmental variation in the genetic model we used, are not likely to strongly affect this trait correlation.
We also found some suggestive evidence that common environmental factors may play a role in forming the association between BMI and education. The correlation between BMI and education was higher than expected among dizygous twins than among monozygous twins if genetic factors alone were to explain these within-pair cross-trait correlations. In the genetic modeling, we found that when the common environmental correlation was included in the model, the fit of the model improved, even though the common environment was not found to have an effect on BMI. An explanation for this paradoxical finding may be that genetic factors due to dominance and common environment both influence BMI, but these effects are not found in univariate analyses when only information about twins reared together is available. In our study, the effect of the common environmental component would be seen as higher dizygous correlations between education and BMI, because dominance effects had no effect on education.
Our finding that genetic factors, and suggestively also common environmental factors, were the main determinants of education-based BMI differences in adulthood agrees with the findings of previous studies and lends further support to the emergence of socioeconomic obesity differences in childhood and adolescence. In a study based on a large cohort of Danish young men, those with BMIs below the median had the highest educational attainment, which declined evenly with increasing BMI. This finding suggested that the association may not have been due to social stigmatization related to obesity but rather to shared common genetic or environmental factors (29). In another Danish study, obesity was found to be associated with school difficulties, which emphasized the role of school age in the development of socioeconomic differences in obesity (33). These results suggest that socioeconomic disparities in obesity start to emerge already before adulthood. In a US study, obese men had lower cognitive function measured after a follow-up of 4–6 y than did nonobese men when education and occupation were adjusted for; this association was not observed in women (34). The authors suggested that this finding indicates a direct effect of obesity on cognitive function through pathophysiologic mechanisms. In light of our results, common genetic factors may also contribute to this association.
The genetic model used in these data assumes random mating. However, our previous study based on the same data disclosed a tendency toward assortative mating by BMI (35). This is partly because of trait assortment, ie, selection of a spouse based on trait, and partly because of the similar background of spouses, ie, social homogamy. There is also evidence of assortative mating by education (36), and this likely exists in our data as well but could not be tested because of the lack of data on spousal education. We are not aware of any previous study on assortative mating for the relation between BMI and education. If a spousal correlation between these 2 factors exists, it should inflate dizygous correlations and further weaken the estimate of the genetic variance component (37). Thus, it is possible that the genetic correlations in our study were underestimated because of the effect of assortative mating.
In this study we analyzed the association between BMI and education, which is a key indicator of socioeconomic status (38). There are several advantages of education as an indicator of socioeconomic status: it is a continuous measure and can be assessed equally in employed and in nonemployed persons; it is a suitable measure in the study of BMI because it shapes behaviors, such as eating and exercise, through knowledge, attitudes, and the values it provides; it is usually completed by early adulthood; and it contributes to occupational status and income. Thus, a later increase in BMI is not likely to affect educational attainment, although it may have an influence on other indicators of socioeconomic status if obesity results in discrimination or downward social mobility. Indeed, previous studies suggest that BMI is associated with unemployment (27) and income level (39) even after education is controlled for.
In conclusion, our results suggest that the origins of education-based differences in obesity are likely to be located in the common background factors of education and obesity and less in the direct association between education and BMI. In light of these results, a reduction in socioeconomic obesity disparities will require interventions starting already in childhood, and such interventions should be targeted at disadvantaged children because later interventions may only have a limited effect.
ACKNOWLEDGMENTS
KS planned the study design, analyzed the data, and drafted the paper. SS-L and EL contributed to the interpretation of the results and sourcing the relevant literature. MK and JK collected the data, assisted in the statistical analyses, and contributed to the interpretations of the results. All authors contributed to the writing of the paper and accepted the final version. The authors had no conflicts of interest.
REFERENCES