Long-term a posteriori dietary patterns and risk of hip fractures in a cohort of women

The Swedish Mammography Cohort (SMC) was established in 1987–1990. All 90,303 women born between 1914 and 1948 and residing in two Swedish counties, were invited to a mammography screening. Enclosed with this invitation was a questionnaire covering diet and lifestyle, which was completed by 74% of the women. Subjects were excluded if the national registration number was missing, the questionnaire had not been dated or energy intakes were deemed as implausible (±3 SD from the mean value of the ln-transformed energy intake). Subjects with prior cancer diagnosis, except those with non-melanoma skin cancer were also excluded. In late 1997 a second expanded questionnaire was sent to those still living in the study area (response rate 70%). The cohort with exclusions has been described previously [24] and consists of 61,433 (baseline) and 38,984 (1997) women. For the present analysis we defined our study sample based on food groups generated with data from the two FFQs in a two-step approach. We excluded individuals with more than 10 of the food groups missing at each investigation as described in Fig. 1, to avoid the poorest assessments but careful not to introduce collider bias by for example influences from socio-economic status [25, 26]. This left 56,736 (baseline) and 35,625 (1997) women for the present analysis. Any remaining missing data on food groups was considered as zero consumption. The study was approved by the regional ethics committees at Uppsala University, Uppsala, and Karolinska Institutet, Stockholm, Sweden.

The dietary assessment [24] (http://​ki.​se/​en/​imm/​the-swedish-mammography-cohort-smc) consisted of food-frequency questionnaires (FFQs) that included 67 and 96 food items in 1987 and 1997, respectively. Participants indicated in the FFQs how often, on average, they had consumed each food item during the past 6 months (1987 FFQ) or the last year (1997 FFQ) and chose from 8 predefined frequency categories ranging from “never/seldom” to “4 or more times per day” (1987 FFQ) or “3 or more times per day” (1997 FFQ). Frequently consumed foods such as dairy products and bread were additionally reported as number of servings per day. Energy and nutrient intakes were estimated by multiplying the consumption frequency of each food item with the nutrient content of age specific portion sizes. Nutrient values were obtained from the Swedish food composition database, National Food Agency. Nutrient intakes were adjusted for total energy intake using the residual method [27].

Frequency data expressed as the consumption frequency per day were collapsed into 31 food groups, based on similarities in nutrients and culinary use. Some food groups consisted of just one food or beverage, for example tea, coffee, eggs and strong beer because they did not fit into any other group or were considered a food group on their own. A previous study conducted in SMC indicated differential effects of fermented milk and non-fermented milk (described onwards as milk) on fracture risk [8]. Therefore, separate milk and fermented milk food groups were formed. Further, since the 1997 FFQ asked about a greater number of food items, partly due to the changing food market, a food group will consequently reflect information from a greater number of food items in 1997 compared to 1987, as indicated in Table 1.

We considered outcomes registered between study baseline (1987–1990) and 31 December 2014. Hip fracture cases were defined by the ICD10 codes S720, S721 and S722. Information on incident hip fracture events was obtained through individual linkage to the Swedish National Patient Registry [28] and by the use of a previously validated and accurate method [29]. The matching was almost complete because of the unique identification number assigned to all Swedish permanent residents.

Covariates were chosen using directed acyclic graphs (DAG) [30]. The covariates obtained from the questionnaires were use of calcium and multivitamin supplements, weight (kg) and height (m), smoking habits, living alone, menopausal status, educational level and leisure time physical activity during the past year. Body Mass Index (BMI) was calculated as weight (kg) divided by height squared (m²). The physical activity question had five pre-defined levels ranging from 1 h to more than 5 h per week and this question has been found valid compared to activity records and accelerometer data [31]. Comorbidity, expressed as Charlson’s weighted comorbidity index [32, 33] was defined by ICD diagnosis codes (versions 8, 9 and 10) from the National Patient Register before baseline and 1 January 1998. Hip fractures that occurred before baseline were indicated as previous fractures. Information on the frequency of falls before baseline and between the two investigations was retrieved from the Swedish National Patient Registry.

Dietary pattern components were formed in exploratory PCA using 31 food groups expressed as the consumption frequency per day, at each investigation separately. Which components to retain from the analysis was primarily based on the Scree plot, eigenvalues (>1.5) and interpretability of the components. The PCA was then repeated with the number of components to retain and components were varimax rotated to become orthogonal (uncorrelated factors). The dietary patterns were interpreted and named on the basis of the components loadings (Table 1). The loadings for each pattern were obtained as a weighted sum of the reported frequencies of food groups and their loadings and each participant was assigned a score for each dietary pattern. To be considered in the Cox-analysis, the dietary pattern had to be formed at both investigations since quartiles of the time-updated dietary patterns were used, using information from both investigations [34]. Correlations between dietary patterns formed at baseline and 1997 as well as between dietary patterns and nutrients were investigated with Spearman rank correlations. For each participant follow-up time was accrued from baseline (primary analysis) or 1 January 1998 (sensitivity analysis) until date of hip fracture, date of death or the end of the study period (31 December 2014) using age as primary time scale. We estimated age and multivariable adjusted hazard ratios by Cox proportional hazards regression and 95% confidence intervals [CI] for the relation between quartiles of time-updated dietary patterns and hip fracture. The proportional hazard assumptions in the Cox models were confirmed graphically by comparing Nelson-Aalen plots and formally with Schoenfeld’s test. The main multivariable adjusted model included height (continuous), educational level (≤9, 10–12, >12 years, other), living alone (yes or no), calcium-supplement (yes or no), multivitamin-use (yes or no), physical activity level (5 categories), previous fractures (yes or no), menopausal-status (yes or no) and Charlson’s comorbidity index (continuous; 1–16) (Model I). In an additional analysis, fermented milk and milk were added to Model I in the analysis with the healthy and Western/convenience patterns, respectively. Further, energy intake, BMI (both continuous) and smoking status (current, former, never)—not primarily considered as confounders according to the DAG—were added to Model I and are referred to as Model II. Continuous covariates were time-updated and missing data on covariates were imputed using Stata’s “mi” package. We used 20 imputations to reduce sampling error. The multiple imputation takes into account model variables and produce 20 separate datasets. The Cox analysis is then run on all separate datasets and the results are combined. The variables living alone, use of calcium and multivitamin supplements, menopausal status, smoking and level of physical activity that were only available in the 1997 questionnaire were imputed, as well as educational level, BMI and height (less than 5% missing). In sensitivity analysis the Cox regression analysis was repeated using complete data, as well as using dietary patterns formed at each investigation. We additionally included history of falls as a covariate in the sensitivity analysis. Further, to study the impact of consistency or change in dietary patterns, the extreme quartiles of the Healthy dietary pattern at each investigation was used to create a three level categorical variable which was coded as [low 1987–low 1997] and set to the reference category, [low 1987–high 1997] and [high 1987–high 1997] and then used in the Cox regression. The baseline of this specific analysis was the second survey (January 1st, 1998). Lastly, study participants were jointly classified in nine strata across the tertiles of the time-updated Healthy and Western/convenience patterns and then used in the Cox regression. P values <0.05 were considered statistically significant. All analyses were carried out in Stata version 12.0 (Stata Corp., College Station, TX, USA).