Prospective associations between diet quality and body mass index in disadvantaged women: the Resilience for Eating and Activity Despite Inequality (READI) study

Study population

Data were provided by women who participated in the Resilience for Eating and Activity Despite Inequality (READI) study; detailed methods are provided elsewhere.³⁴ READI is a cohort study of women living in socioeconomically disadvantaged areas within Victoria, Australia, and was designed to investigate pathways by which socioeconomic disadvantage influences behaviours associated with obesity risk. The study was approved by the Deakin University Human Research Ethics Committee. Women provided written, informed consent to participate.

The Index of Relative Socioeconomic Disadvantage³⁵ was used to classify all neighbourhoods (suburbs) in Victoria, Australia, within urban/rural strata into tertiles of disadvantage. Within the most disadvantaged tertile, 40 urban and 40 rural neighbourhoods were randomly selected. The electoral roll (voting is compulsory in Australia) was then used to randomly identify 150 women aged 18–46 years from each of these neighbourhoods (n = 11 940, as some included areas had fewer than 150 eligible women; n = 11 079 surveys were able to be delivered), who were invited to complete a mailed questionnaire on sociodemographic, anthropometric and lifestyle-related variables.

A total of 4938 women (45% response rate) completed a survey at baseline, of whom 4349 were eligible to participate (39% of those delivered a survey) (2007/08; Figure 1). At baseline, non-respondents were more likely to reside in an urban than a rural area, and resided in areas with greater area-level disadvantage compared with respondents (P < 0.01).³⁴ Compared with the general population of women living in the 80 sampled neighbourhoods, a greater proportion of READI women at baseline were Australian-born (89% vs 73%) and were married or living as married (65% vs 49%), whereas a lower proportion were in full-time employment (37% vs 58%).³⁴ In 2010/11, 1913 women completed surveys, with 1560 (36% of initial cohort) completing surveys in 2012/13. Bias may be introduced when individuals lost to follow-up differ in both exposure status and outcome compared with those who remain. In this cohort, although those who were lost to follow-up had a lower mean Dietary Guideline Index (DGI) score at baseline (78.8 vs 82.2, P < 0.0001), mean BMI did not differ among women who were and were not lost to follow-up. The final analytical sample (n = 1242) represented 80% of eligible women who completed a survey at all three time points. As described in Figure 1, primary reasons for exclusion included pregnancy, poor self-rated health and missing ≥ 50% of BMI values or questionnaire responses.

Questionnaire

Multiple imputation

Of the 1242 eligible women, 887 (71.4%) had complete data for all variables. Among the variables considered in the analysis, the proportion of missing values across all time points was 2.3% for BMI, 2.1% for IPAQ subscores and < 1% for DGI subscores and for smoking status, educational level, marital status and number of births. There were no missing values for age, menopausal status or urban/rural residence. To reduce bias due to item non-response, missing data were imputed using chained equations multiple imputation with the user-written ice command^45,46 in Stata (version 12.0; Stata Corp, TX, USA) under a missing at random assumption. Imputation models included BMI, age, smoking status, menopausal status, educational level, marital status and number of births at each of the three time points, and urban/rural residence at baseline. In addition, the 12 DGI subscores and the four IPAQ subscores were used in the imputation models, rather than the total scores, to retain as much complete information as possible. Thirty imputations were performed to correspond to the proportion of participants with missing data.⁴⁷ Imputed BMI values (n = 86) were deleted, as removing imputed outcomes provides more precise estimates and can minimize bias introduced by a mis-specified model.⁴⁸

Statistical analyses

Associations between concurrent change in diet quality (DGI score, mean-centred, continuous) and BMI (continuous) were assessed over the three time points in Stata (version 12.0; Stata Corp, TX, USA). Concurrent changes were considered, as they yield more robust, consistent and biologically plausible results compared with analyses based on prevalent or lagged-change in diet.⁴⁹

Fixed effects and mixed models allow estimation of subject-specific associations in longitudinal studies. As each approach has distinct capabilities, both models were fitted in order to understand how their effect estimates might differ. The first model was a fixed effects model with robust standard errors. This model examined within-individual change accounting for any time-invariant confounding (measured and unmeasured), since each woman acts as her own control. Importantly, only observations where the exposure changes contribute to their estimates.⁵⁰ In this model, the coefficient of DGI represents the average change in BMI for a 1 unit within-individual change in DGI score.

The second model was a mixed model with level 1 (measurement occasion) and level 2 (individuals) random effects, and clustered standard errors at the neighbourhood level. The intercept and slope for age were allowed to vary randomly, meaning each woman had her own trajectory of change in BMI. Models were fitted using maximum likelihood estimation with an unstructured correlation matrix. Although their effect estimates may be biased by unmeasured time-invariant confounders, mixed models provide more precise estimates, permit explicit modelling and partitioning of variance structures both within and between individuals, and can estimate effects for time-invariant factors.^50–52 The coefficients for time-varying predictors in mixed models are a combination of between- (i.e. the average difference in BMI associated with a 1-unit increase in DGI score between women) and within-individual effects (i.e. the average change in BMI associated with a 1-unit increase in DGI score within a woman).

The primary analyses examined findings in the imputed data set (n = 1242 participants). All models adjusted for the following time-variant confounders: age (mean centred), smoking status, menopausal status, educational level, marital status, number of births and physical activity. Urban/rural location of residence at baseline was included as a time-invariant confounder in the mixed model only. The interaction between age and DGI score was included to determine if the change in BMI with age differed depending on DGI score.

Sensitivity analyses

In sensitivity analyses, a complete case analysis (n = 887) was conducted under the assumption that the data were missing completely at random.

Characteristics of study cohort

At baseline, women were on average 36.9 [standard deviation (SD) = 7.5] years of age. The majority were living with a partner (71.9%), had children (66.9%) and had completed secondary school and/or a trade/diploma (46.6%) (Table 2). Mean BMI increased from 25.8 kg/m² (SD = 5.7) at baseline, to 26.5 kg/m² (SD = 6.1) at wave 3, and the mean DGI score increased from 83.1 (SD = 18.3) at baseline to 85.4 (SD = 16.9) at wave 3.

Associations between change in DGI score and change in BMI

Table 3 shows the estimates from the two different modelling approaches, with and without the interaction between age and DGI score included in the model. Results from the fixed effects model indicated that the average BMI for a given woman increased by 0.14 kg/m² [95% confidence interval (CI): 0.11, 0.17, P < 0.0001] with each year increase in age. Similar results were obtained from the mixed model (increase of 0.13 kg/m², 95% CI: 0.10, 0.16, P < 0.0001). From the fixed effects model that included the interaction term, average BMI decreased by 0.014 kg/m² (95% CI: -0.021, -0.007, P < 0.0001) for a woman of average age (37 years) with each unit increase in DGI score. The mixed model coefficient, representing a combination of within- and between-individual effects, was estimated to be -0.012 kg/m² (95% CI: -0.019, -0.005, P = 0.001). The negative coefficient for the interaction between age and DGI score suggests that the rate of change in BMI with age was greater for those with a lower DGI score than for those with a higher DGI score (fixed effects model P = 0.09, mixed model P = 0.07; Figure 2).

Sensitivity analyses

The descriptive baseline characteristics of the complete cases were similar to those of the imputed sample, and overall findings were unchanged in the complete case analysis (see Supplementary file, available as Supplementary data at IJE online). However, although the effect estimates for the interaction between age and DGI score were similar in analyses based on complete cases and those with imputed data, the P-values and confidence intervals were larger in the complete case analysis (fixed effects model P = 0.22, mixed model P = 0.16), potentially due to the smaller sample size.

Strengths and limitations

This study was strengthened by use of a validated food-based diet quality score which helps to minimize confounding due to biological and metabolic interactions among nutrients and foods. Moreover, we compared estimates from two powerful participant-specific modelling techniques, assessed diet quality, potential confounders and BMI at three time points, and used multiple imputation to reduce bias associated with item non-response.

The primary limitation of this study is that measures of BMI and food intake were self-reported, and both tend to be under-reported among women.^58,59 However, the FFQ was adapted from validated measures, and self-reported BMI corresponds closely with measured BMI in Australian women.⁶⁰ Moreover, under-reporting would tend to attenuate associations, and if women under-reported similarly at all time points there would be little to no bias in the change value. Data were adjusted for physical activity level, which was also assessed on the basis of self-reported data. Although the IPAQ has been shown to have acceptable reliability and validity,^43,44 residual confounding may nevertheless exist due to mis-reporting. As the current study used a brief FFQ, we were unable to adjust for total energy intake, although we did adjust for age and physical activity level which are major determinants of energy intake. In addition, many similar studies have shown that adjustment for energy intakes has little to no impact on estimates,^{24,28,30,56,61,62} whereas others contend that energy adjustment is not required as energy intake may mediate the effects of diet quality on weight gain.^62,63

It is possible that women who volunteered to participate in the study were more health-conscious than those who chose not to participate, or that participation caused them to become more health-conscious. However, mean BMI and DGI scores were poor at all three waves, and were similar to values reported in large national surveys.^57,64 We also updated DGI scores and BMI at each wave, and adjusted for other time-variant behaviours that could influence diet quality and BMI, such as smoking and physical activity. Provided that lifestyle behaviours and BMI were accurately reported, or reported with a similar degree of error over time, and that important confounders were not omitted, our estimates of change should be unbiased. A further limitation is that we were unable to consider neighbourhood of residence as a potential time-varying confounder.

Factors that might relate to missingness were included in imputation and analytical models to make the missing at random assumption more plausible. Scale scores were used during imputation, as it was computationally infeasible to include all of the individual dietary and physical activity items, and because imputing only the total DGI and IPAQ scores would have discarded real values from subscores with complete data. Imputing scale subscores has been shown to be superior to complete case analyses and to imputing total scale scores.^65,66 Because we imputed DGI subscores rather than the total score, our imputation models did not contain the total DGI score, nor the interaction between age and DGI score. However, DGI subscores should be highly correlated with the total score, and the interaction between age and DGI score was very small.

Implications for policy and research

Nutrient- and energy-focussed metrics are increasingly being challenged by evidence indicating their inadequacy in explaining patterns of chronic disease and weight gain, and have therefore been superseded by a focus on consumption of whole foods and overall dietary patterns.^67,68 Acknowledging the self-reported nature of the data, the current study supports this paradigm shift and demonstrates that adherence to food-based recommendations can assist women living in disadvantaged neighbourhoods to minimize weight gain. Such recommendations are likely to be more readily understood and acted upon than nutrient-based recommendations.⁶⁹ As research proceeds to identify particular foods and combinations thereof that can help to support maintenance of a healthy body weight, researchers should be careful to apply robust modelling strategies that maximize the power of their analyses in order to generate more robust evidence from which to formulate dietary recommendations.

Conclusions

Positive change in diet quality consistent with the recommendations in the Dietary Guidelines for Australians is associated with reduced BMI gain among women living in disadvantaged neighbourhoods. Although small, these changes are important at a population level. Thus, supporting disadvantaged women to adhere to population-level dietary recommendations is likely to assist them with long-term weight management.