Education differences in women’s body weight

Empirical evidence from the past four decades of research has shown that body weight is strongly and negatively associated with socioeconomic status in developed societies [14,16]. A consistent finding is that this association is stronger among women [1]. For example, findings from the U.S., France, and Germany showed that the body weight gap between the top and the bottom tiers of education was three times higher among middle-aged women than among middle-aged men [17].

Studies based on longitudinal and repeated cross-sectional data have shown that education gaps in body weight emerged in early adulthood (approximately during the age between 20 and 30) and still widened until middle adulthood (approximately until the age of 50) [7,8,17,18]. For example, Ailshire and House [7] examined panel data from the Americans’ Changing Lives Survey (ACLS) and found that differences in body weight between lower educated black women and higher educated white women grew especially between the ages of 25 and 54 (by 0.1 of Body Mass Index (BMI) points annually).

Similar tendencies were found in other countries. Molarius and colleagues [17] used repeated cross-sectional data from the comparative “Monitoring Trends and Determinants in Cardiovascular Disease” project (MONICA) of the World Health Organization. Their results revealed that in most developed countries, education differences in BMI grew especially among women below the age of 44. The average increase across ten years ranged between +0.1 BMI points in Italy and Belgium and +1.3 BMI points in Germany and Poland [17]. This evidence suggests that education differences in body weight among women typically increase during a life stage ranging approximately from the early twenties to the mid-forties.

The magnitude of divergence is alarming, given that overweight and obesity are associated with a range of negative health-related, social, psychological, and economic outcomes [2,3]. Earlier onset of overweight and obesity among lower educated women, might mean longer duration of exposure and more severe consequences in terms of physical and mental health [4]. Moreover, overweight and obese people (and especially women) experience disadvantages on the labor market [6], which may be most consequential in early career stages. Finally, even women who are not overweight show lower subjective well-being, higher risk of depression, and lower self-esteem when they gain weight [19]. Understanding the mechanisms behind education differences in weight gain is therefore relevant to the study of inequality in various domains of life.

Education differences in women’s weight trajectories: The role of motherhood

What explains diverging weight trajectories between education groups throughout early and middle adulthood? Although the phenomenon is well documented, little is known about the underlying mechanisms [8,17,18]. According to the life-course perspective [20], events and transitions experienced early in life represent turning points that channel individuals into adverse trajectories (“chains of risks”) or beneficial trajectories (“protective chains”). A general expectation of the cumulative (dis)advantage theory [21–23] is that the incidence and timing as well as the direction and magnitude of the effects of major events and transitions are socially stratified. As a result, groups that initially lack socioeconomic resources are channeled into adverse trajectories, whereas advantaged groups enter beneficial trajectories, leading to an increase of social inequality over the life course [3,24,25].

Motherhood may represent a turning point for education differences in body weight over the life course. This would imply that motherhood (a) is related to weight gain, (b) differs by education in terms of prevalence and/or timing, and/or (c) differs by education in terms of vulnerability to its effects on body weight. As we explain in the following sections, previous research has offered empirical evidence in support of each of these arguments.

Motherhood as a risk factor for weight gain

Education, motherhood and weight gain in the German context

If motherhood is a turning point that triggers the increase of education differences in women’s body weight, its prevalence, timing, and/or the vulnerability to its effects on body weight are expected to differ between education groups. Previous research provides evidence in support of each of these arguments in the German context of this study.

First, the prevalence of motherhood differs between education groups in Germany [32,33]. Lower educated German mothers had 2.5 children on average and 27% had three or more children; higher educated mothers had 1.9 children on average and only 12% had three or more children [34].

Second, education also structures the timing of motherhood in Germany, as most women follow the “sequencing norm” of finishing education before having children [35,36]. Among higher educated German women born between 1952 and 1972, the median age at first birth was 33; among lower educated women of the same cohorts, the median age at first birth was 26 [11].

Third, vulnerability to the effects of motherhood on body weight is likely to differ between education groups. Although no study has examined education differences in the effects of motherhood on body weight in Germany, recent research from other Western countries has shown that lower educated women are at a higher risk of excessive weight gain during pregnancy and retain more weight post-partum than higher educated women [12,13,37–39]. For example, in the U.K., lower educated women retained 3.2 kg eight months post-partum, compared to only 1.8 kg among higher educated women [38].

Moreover, the proposed mechanisms behind differential vulnerability to the effects of motherhood on body weight are also likely to operate in the German context. First, the direct effects of pregnancy and childbirth were found to be stronger for women with menarche at younger ages (<12 year old) and for women who gave birth between the age of 20 and 30 (compared to women who became mothers after the age of 30). Although these findings were not linked to education differences in the effects of motherhood, other studies have found that lower educated women experienced their menarche at younger ages [5] and also gave birth at younger ages [11].

Second, compared to higher educated mothers, lower educated mothers more often lack social, economic, and personal resources that help cope with parenting stress [40,41], which in turn may strengthen the chain of risk factors resulting in weight gain. In Germany, protective resources are strongly stratified by education [42,43], suggesting that lower educated women are more vulnerable to the effects of motherhood on weight gain.

Taken together, these considerations and the supporting empirical evidence suggest that motherhood may constitute a turning point in the life-course trajectory of education differences in body weight. Although this idea has been articulated in previous studies [7,44], no research has examined the extent to which the increase of education differences in body weight during reproductive age is explained by differences in the prevalence, timing, and/or effects of motherhood. Previous studies aiming to explain education differences in body weight did not examine the role of these factors from a life-course perspective, particularly in relation to the distinctive pattern of increasing gaps from early to middle adulthood [15].

Hypotheses

Based on these theoretical considerations and previous evidence, we formulate the following set of hypotheses to guide our empirical analyses: Education differences in women’s body weight emerge and increase during reproductive age (Hypothesis 1). Transitions to motherhood and higher parities are associated with an increase in body weight (Hypothesis 2). Motherhood is associated with larger weight gains for lower educated women than for higher educated women (Hypothesis 3). Education differences in trajectories of body weight are explained by education differences in terms of (a) the prevalence and timing of motherhood, and (b) vulnerability to the effects of motherhood on weight gain (Hypothesis 4).

Sample selection

Our analyses were based on data from the German Socio-Economic Panel Study [45] v.34. The SOEP is a representative household survey with annual interviews; it started in 1984. Currently it includes data on more than 20,000 respondents from more than 10,000 households drawn from the original sample as well as from regular refreshment samples [46]. The SOEP is among the largest and longest-running representative panel surveys that exist and is widely recognized for maintaining the highest standards of data quality and research ethics. Information on respondents’ body weight and height has been collected in the SOEP since 2002 in biannual intervals. Our analysis drew on these data from an observation period between 2002 and 2016.

In 2002, the anchor year of our study, the sample comprised 29,101 individuals. For the purposes of our investigation, we excluded men (n = 14,244) and respondents outside the age range between 17 and 45, as well as those who were older than 35 at first observation (n = 11,397). These restrictions centered the analysis on women’s reproductive period and allowed us to examine changes over this life stage. We also excluded immigrant women who were older than 15 years when they arrived in Germany (n = 206) to ensure that all women in our sample completed their education in Germany.

Moreover, we excluded observations with extreme values of BMI (lower than 16 or higher than 59) to minimize outlier effects (n = 4) and we excluded all observations with missing data on one of the covariates (n = 199). Finally, we excluded individuals (n = 384) who were observed only once, as our statistical models required at least two observations per respondent. The final sample consisted of 2,668 women, comprising 13,899 panel observations. All sample exclusions are summarized in Table A1 in the S1 Appendix.

Measure of education

Education was measured as a time-constant indicator of the highest degree attained until the end of the observation window. Our observation window of up to 14 years covered the age span during which even the youngest respondents have most commonly received their highest educational degrees, ensuring that our measure accurately reflected education level.

Our education measure was based on the “Comparative Analysis of Social Mobility in Industrial Nations” classification (CASMIN). We grouped the CASMIN categories as follows: The bottom category comprised individuals holding lower secondary degrees with completed vocational qualification or less (CASMIN 1a–1c); intermediate education ranged from intermediate secondary degrees to higher secondary degrees with vocational qualification (CASMIN 2a–2c); the top category included respondents holding tertiary degrees (CASMIN 3a–3b). In our sample, 24.2% of women were lower educated, 54.2% had intermediate degrees, and 21.6% were higher educated (see Table 1). The share of missing information on education amounted to only 2.9%. Cases with missing information on education were removed from the analyses (see Table A1 in the S1 Appendix). In substantive terms, these categories represent meaningful comparison groups for the study of the role of motherhood in education differences in body weight, as the prevalence and the timing of motherhood–as well as resources that may help to prevent weight gain due to motherhood–differ substantially between these groups.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics.

https://doi.org/10.1371/journal.pone.0236487.t001

First, the time spent in the education system strongly determines the timing of motherhood in Germany. Due to a strong sequencing norm, most women in Germany become mothers after completing their education (Fig 1) [11]. As visible from Table 1, lower educated mothers’ average age at first birth was 24, approximately 6 years less than higher educated mothers’ average age at first birth.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Timing of motherhood by education.

Data are from SOEP, v.34; own calculation.

https://doi.org/10.1371/journal.pone.0236487.g001

Second, social and economic resources are strongly stratified by education in Germany, which is mainly due to a selective and rigid school system characterized by early tracking and a strong connection between educational degrees and labor market outcomes.

These conditions favor the reproduction of initial advantages and disadvantages related to social origin, and stratify economic outcomes in later life along educational lines [47,48]. These features of the German context suggest that lower educated women are more vulnerable to the effects of motherhood because they are often deprived of resources protecting them from the chain of risks resulting in weight gain.

Measures of body weight

Our outcome measure was body weight in kilograms (controlled for self-reported height). This measure was based on self-reported information on respondents’ weight and height, assessed biannually since 2002. Self-reported weight and height measures were shown to be closely related to physical measurements. While people report their height relatively accurately, the discrepancy between measured and self-reported weight is larger and related to age, education, sex, and body weight itself. Specifically, younger, higher-educated, and rather heavy women tend to underestimate their weight [49–51]. However, the size of these discrepancies is relatively small as about 85% of self-reports of weight correspond closely with measured weight [52]. For these reasons, self-reported weight and height are considered valid measures in studies of population health, although some caution is warranted with the interpretation of effect sizes. In our models, we focus on body weight (controlled for self-reported height) to ease the interpretation of effect sizes. As BMI provides important additional information regarding the thresholds of overweight and/or obesity we also presents results for BMI. Table 1 provides descriptive information on how body weight and height varied between education groups. Averaged across all panel observations, lower educated women were smaller but 2kg heavier than women with intermediate degrees and 5kg heavier than women with higher degrees.

Measures of motherhood

In studies on body weight, motherhood is typically measured by indicators for the number of children [9,31]. In our analysis, we complemented this standard specification by a measure of short-term effects of motherhood related to pregnancy and the time after birth. The first measure for the number of children was a time-varying categorical variable distinguishing between observations in which women were childless (reference category), had one child, had two children, or had three or more children. The second measure capturing the short-term effects of pregnancy and motherhood on body weight was a dummy variable coded 1 if women gave birth in the year of the interview or in the previous year and 0 otherwise. A number of studies have suggested that the effects of motherhood might be particularly strong at higher parities. To examine this possibility, we additionally assessed (1) whether the short-term effects were larger at higher parities, and (2) whether the short-term effects at higher parities differed by education. The results showed that the short-term effect of motherhood was only slightly larger at parities of 2 and above. It was also slightly larger among lower educated women. Because all substantive conclusions remained the same under this alternative specification, we decided to keep the model parsimonious and did not include further interactions.

Moreover, as being pregnant in the reference year might lead to an underestimation of the effects of motherhood, we assessed whether our results were robust to excluding observations during pregnancy. This exclusion did not influence on our results. For the main analyses, we kept these observations in the data. Table 1 shows large education differences in the motherhood indicators. Lower educated women were childless in only 32% of the observations, compared to 42% for intermediately educated women, and 61% for higher educated women. Large education differences were also observed for higher parities. Lower educated women had three or more children in 19% of their observations, intermediately educated women in 8% of their observations, and higher educated women in 4% of their observations. These descriptive results were in line with previous research [34] and our theoretical expectations about education differences in motherhood.

Fig 2 further shows that lower educated women became mothers earlier in life than higher educated women. Higher educated women had ‘caught up’ on having one child (by the age of 30) or two children (by the age of 38), but not on having at least three children. This category was much more prevalent among lower educated women across the entire age range of our study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Distribution of the number of children categories by education and age.

Data are from SOEP, v.34; own calculations; Lower education = dashed lines; Higher education = solid lines.

https://doi.org/10.1371/journal.pone.0236487.g002

Modelling of age effects

Age was measured in years and months at the time of the interview. The age range covered was 16.5 to 45.6 years, with a mean of 31.6 years. In the models (Table 2), we included linear and squared terms of age, centered at the age of 30.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Panel models for body weight.

https://doi.org/10.1371/journal.pone.0236487.t002

We tested for alternative specifications of the functional form of age effects, such as linear, squared, and cubic terms of age as well as the most flexible non-parametric specification with dummies for each year of age. According to the Bayesian Information Criterion (BIC) and diagnostic plots, a model that included linear and squared terms of age and their interactions with education provided the best fit to the data. Fig 3 shows a comparison between our preferred model for body weight and a semi-parametric model estimating body weight differences separately for each age. The figure indicates that our parametric model accurately described body weight trajectories for different educational groups.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Predicted body weight trajectories form the preferred parametric model specification and from a semi-parametric model specification.

Data are from SOEP, v.34; own calculations; (a) predicted body-weight trajectories form the preferred parametric model specification (model M1, Table 2); (b) predicted body-weight trajectories form a semi-parametric model specification; 95% CI = grey areas.

https://doi.org/10.1371/journal.pone.0236487.g003

Given the structure of our data, in which individuals enter the data at different ages and are observed for a maximum of 15 years, cohort effects might be confounded with age patterns. To assess the influence of cohort effects, we examined interactions between cohort and age (linear, squared, cubic) and education. According to the BIC criterion and diagnostic plots, these additional interactions worsened the model fit. Therefore, we did not control for cohort in our analyses. This means that our analysis assessed the relationship between education, age, and body weight averaged over all cohorts.

Weighting and attrition

We used cross-sectional survey weights to ensure that our data were representative of the German population. These weights correct for oversamples included in the SOEP as well as systematic panel attrition before our starting year of 2002. The survey weights were calculated by the SOEP research team [53]. After weighting, our sample was representative for women aged 17 to 45 living in German households in 2002 [54].

Furthermore, we analyzed whether panel dropout was related to our variables of interest, estimating a discrete-time survival model for the probability of dropout, defined as providing no or no valid information before the end of the observation period in 2016. Our results indicated that dropout was barely related to motherhood, education, and body weight. Mothers were slightly more likely to stay in the panel and this did not vary by education or body weight. Higher educated women were slightly more likely to drop out when they gained weight, while lower educated women were slightly more likely to drop out when they lost weight. Although this suggested that educational differences in body weight might be overestimated if selective dropout is ignored, this bias is very small.

Analytic strategy

The aim of our analyses was to estimate age-related change of education differences in body weight and to assess the role of motherhood in this process.

To estimate the effects of motherhood, we used fixed-effects models, which account for time-stable characteristics of women that may confound the effect of motherhood. The fixed-effects estimator draws only on within-person change over time, controlling for all time-constant characteristics of women, regardless of whether they are measured or not.

In our case this means that the effects of motherhood are estimated solely from actual transitions to motherhood and to higher parities. Body weight is observed before pregnancy and changes are observed during pregnancy and after giving birth. The main advantage of this approach is that time-constant confounders are rendered inconsequential because the estimates are based solely on change within persons over time.

To estimate initial educational differences in body weight, we used a between-person estimator, given that a fixed-effects model does not estimate the effects of a time-constant variable such as women’s education. To obtain a between-person estimator, we used OLS to calculate initial educational differences in body weight at age 19.

This means that our models combined within-person fixed-effects estimators with an initial between-person OLS estimator [55]. This approach is similar to the hybrid model [56,57] and provides within-person fixed-effects estimates for the effects of age and motherhood (i.e., within-effects) while allowing for the inclusion of time-constant variables such as education and height to estimate initial differences in body weight (i.e., between-effects).

Based on this approach, we estimated three models to answer the main questions of our study. The first model described the phenomenon we sought to explain, namely education differences in body weight across the reproductive period of women’s life courses (Hypothesis 1). The second model assessed the relationship between motherhood and body weight (Hypothesis 2). The third model examined education differences in the effects of motherhood and the extent to which motherhood, and differences in its effects, explained the life-course pattern of education differences in body weight (Hypothesis 3).

We specified our first model as follows: (M1) where Y_it is the outcome (body weight) for an individual i at time point t. γ₁ and γ₂ are the parameters of linear (z_it) and squared () terms of age defined for the reference group of lower educated women (k = 0). γ_3,k and γ_4,k are the differences in linear and quadratic body weight trajectories between the lower, intermediate, and higher educated women. is a time-constant indicator of education. u_i represents the individual-specific time-constant intercept (including all time constant confounders). ϵ_it is the idiosyncratic observation-specific error term with an expectation of zero and assumed to be uncorrelated with all other terms in the model. ∑_{k∈{int,high}} δ_k indicates initial differences in body weight between education groups.

The fixed-effects estimator used to estimate the effect of age on change in body weight was specified as follows: (1A) where indicates within-person change in body weight, calculated as the body weight of individual i at time point t minus the average body weight of an individual i, and () and represent within-person changes in age and age squared; u_i is eliminated from the model. Although this approach solves the problem of time-constant unobserved confounding, it also eliminates the time-constant average level of weight represented by α_t0 and the respective initial differences between the educational groups represented by in Eq 1. Therefore, we added an estimation step for initial educational differences at the age of 19.

(1B)

For the estimation of the initial differences (represented by Y_i19), we additionally controlled for self-reported height () to account for structural differences in height between education groups. The results of these models are presented in Table 2 (M1) and in Fig 3.

In our second model, we assessed how motherhood was related to body weight. For this, we added both indicators of motherhood (M_it)–the number of children and the indicator for whether the respondent gave birth in the year of the interview or in the previous year.

(M2)

The parameters of the model are also estimated using the within-person fixed-effects estimator [55] specified as follows: (2) where represent within-person changes in the motherhood indicators. The results of this model are presented in Table 2 (M2).

In our final model, we further added interactions between education and both indicators of motherhood to assess whether the effects of motherhood on body weight differed between educational groups. (M3) where . Educational differences in the effects of motherhood on body weight are represented by , which allows the coefficients of the birth variables to vary across educational groups. ζ_low represents the estimated body weight of the reference group of lower educated women. The model is estimated analogously to (2), but allowing for differential effects of motherhood as specified. The results of this model are presented in Table 2 (M3).

Combining the results from models 1b and 2 gives us the model-based predictions for weight trajectories of different educational groups.

(3)

In order to answer our main research question–to what extent differences in the prevalence and effects of motherhood account for educational differences in trajectories of body weight (Hypothesis 4)–we estimated a counterfactual scenario based on Eq 3. In this scenario, we estimated, first, what the age pattern of change in body weight among lower educated women would look like if they had the same prevalence of motherhood as higher educated women. For our sample, this means that we predicted how body weight among lower educated women would change with age if they had no children in 61% instead of 32% of their observations, if they had one or two children in 17% instead of 23% of their observations, if they had at least three children in 4% instead of 19% of their observations, and if they had given birth in 11% instead of 9% of their interview years. Second, we used the estimated coefficients of both indicators of motherhood on the body weight of higher educated women () for the calculation of the counterfactually predicted weight trajectory of lower educated women. This allowed us to additionally account for education differences in the effects of motherhood on body weight.

Using these two counterfactual assumptions, the implied counterfactual differences in weight () that remain are defined as: (4) for the scenario in which lower educated women had the same prevalence of motherhood as higher educated women. For both prevalence and effects, the difference is defined as: (5)

As in the equations above, z represents age, M represents the motherhood indicators, and ζ represents the association between motherhood and education.

As educational differences in our outcomes are expected to change with age, it is also likely that the contribution of motherhood–both with respect to prevalence and to effects–to these differences is not uniform across the age range. To account for this, we calculated the contribution of motherhood according to the counterfactual scenario at different ages (z). The results of the counterfactual analysis are presented in Fig 4 and in Table 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Predicted differences in body weight between lower and higher educated women with and without adjustment for motherhood.

SOEP data, v.34; own calculations; height is fixed at the average of 167 cm in both models; black solid reference line = body weight of higher educated women, black dashed line = difference in body weight between lower educated women and higher educated women based on model M1, Table 2; grey dashed line = difference in body weight between lower educated women and higher educated women adjusted for differences in prevalence of motherhood; grey solid line = difference in body weight between lower educated women and higher educated women adjusted for differences in prevalence and effects of motherhood (M3), Table 2.

https://doi.org/10.1371/journal.pone.0236487.g004