The STORK Groruddalen study is a population-based cohort study of 823 healthy pregnant women attending the Child Health Clinics (CHC) for antenatal care in three administrative city districts in the area of Groruddalen, Oslo, Norway, May 2008–May 2010 (22). The study and its methods have been described in detail elsewhere (22). Information about the study was widely distributed in the three city districts of Groruddalen prior to the start of the study. General practitioners were asked to refer pregnant women to the CHC as early as possible. Midwives and research staff recruited the women at their first visit to the CHC in early pregnancy. All information material and questionnaires were translated into eight languages (22) and quality controlled by bilingual health professionals. Professional interpreters were used when needed. Women were eligible if they: (1) lived in one of the three city districts of Groruddalen; (2) planned to give birth at one of two study hospitals; (3) were less than 20 weeks pregnant at inclusion; (4) could communicate in Norwegian or any of the eight translated languages; and (5) were able to give a written consent to participate. Women with pregestational diabetes or other diseases necessitating intensive hospital follow-up during pregnancy were excluded. The participation rate was 74% (64–83% in main ethnic groups) and 59% of the participants were of non-Western origin [for flow chart, see method paper (22)]. The sample was found representative for the population of pregnant women and the major ethnic minority groups living in the Groruddalen area who attend the CHCs (75–85% of the pregnant population) (22, 23). The study was approved by The Regional Ethics Committee and The Norwegian Data Inspectorate.

Information about ethnic origin was collected in gestational week 12±2, on average (visit 1). Ethnicity was defined as the country of birth if ethnic Norwegian, first-generation immigrant or second-generation European immigrant, and the participant's mother's country of birth if second-generation non-European immigrant (only 6.6% (n=50) of the sample were second-generation immigrants) (24). Due to small numbers in some groups and to retain power in the statistical analyses, ethnic origin was merged into the region categories ‘Europe’, ‘South and East Asia’, and ‘The Middle East and Africa’. The category ‘Europe’ (n=352) consisted mainly of ethnic Norwegians (83%) and Eastern Europeans (11.6%). The remaining 5.4% were of Nordic, other European, or Western origin. The category ‘South and East Asia’ (n=231) consisted mainly of Pakistanis (52.4%), 24.7% from Sri Lanka, and 17.3% from East Asia. The remaining 5.6% were of other South Asian origin. The ‘Middle East and Africa’ category (n=174) consisted of women from Iraq (20.1%), Turkey (14.4%), Morocco (13.2%), Africa (43.1%), and the remaining 9.2% from Central Asia (n=13) and South and Central America (n=10).

Socio-economic and integration variables were collected at visit 1. Less than 1% missing values were detected on the variables concerning socio-economic status and integration level, except in the variables: number of rooms in household (182 missing; 22%) and housing tenure (47 missing; 6%). This was due to the questionnaire structure as the study personnel sometimes forgot this question. Level of missing data was similar across ethnic groups, maternal educational level and study personnel, indicating that the missing values were random. In order to replace missing values in the final data set, maternal and household socioeconomic markers were used as both predictors and dependents during a multiple imputation, generating five imputations using linear and logistic regression models. The differences between these five complete data sets, and between actual and imputed data, were minimal. One of the five imputed data sets was chosen by simple randomisation, as running pooled analyses into the principal component analysis was considered inappropriate. Ethnic Norwegians and Nordic participants were not asked questions regarding integration. Hence, they were given top scores for all integration variables.

A principal component analysis was performed on all collected socio-economic and integration-related variables (15 variables) (22). Varimax rotation with Kaiser normalisation was used to produce indicators of socio-economic status and integration level. Four variables (house type, number of rooms in household, ownership of car, and marital status) were excluded from the final analysis because of low factor loadings (<0.4) and low correlation with other socio-economic variables. All variables were graded in the same direction from low to high socio-economic status or level of integration, respectively. Two components, socio-economic score and integration score, explaining 56% of the variance, were recognised in the material in agreement with the scree plot and Kaiser's criterion. Both components had high reliability with Cronbach's α >0.7. Factor scores of each participant were saved through the regression method.

The socio-economic component was mainly defined by five variables with factor loadings ranging from 0.705 to 0.584. The variables were, in decreasing order: occupation (using the International Standard Classification of Occupations (ISCO-08); educational level; tenure (defined as owning vs. renting); level of household crowding (defined as persons in household per room); and employment status (defined as not paid work vs. paid work outside home).

Variables indicating integration level were based on questions from The Oslo Immigrant Health Study (25).The integration component was mainly defined by seven variables with factor loadings ranging from 0.867 to 0.484. The variables were, in decreasing order: self-reported proficiency in the Norwegian language; need of interpreter at doctor's appointments; need of interpreter during the study interviews; duration of residence; how often visited (in any context) by an ethnic Norwegian; frequency of reading Norwegian newspapers or watching Norwegian TV; and occupation.

The individual factor scores for the socio-economic score variable ranged from −2.9 (indicating the lowest socio-economic position) to 2.6 (indicating the highest socio-economic position). The individual factor scores for the integration score variable ranged from −3.6 (indicating the lowest integration level) to 1.6 (indicating the highest integration level).

Habitual diet the previous 2 weeks was characterised using a food frequency questionnaire (FFQ) in gestational week 28±2 (visit 2). The semi-quantitative FFQ was administered by trained midwives. The FFQ was created by nutritionists for this study with combined experience in developing FFQs and knowledge of dietary habits in ethnic minority groups. The FFQ was designed to capture the frequency of intake for food items considered to modify the risk of T2DM and obesity: sugary drinks (26); rapidly absorbable carbohydrates (27) (sweet cakes, white bread, etc.); whole-grain (28); fruit and vegetables (29); beans and lentils (30); dietary fats (31) (lean vs. fatty fish and meat, fatty cakes, etc.). Frequencies of intake were given for 67 food and beverage items. The food items were fruits (1 item); vegetables (2 items); legumes (2 items); potatoes (3 items); the categories bread, cereals, pasta, rice, couscous, and other staples (6 items); yoghurt with or without added sugar (4 items); meat (4 items); fish (5 items); spreads for sandwiches (11 items); cakes, desserts, and confectionery (10 items); salty snacks (3 items); and beverages and sugar added to coffee or tea (16 items). Two questions captured between-meal snacks. Frequency of intake was captured by using either five or seven frequency intervals ranging from ‘never’ to ‘several times weekly’ or ‘never’ to ‘daily’, respectively. Portion size estimates were given for beverages and added sugar to tea or coffee only (16 items). Sugar from beverages was calculated by multiplying the volume of each beverage by mean values of sugar content (based on the Norwegian food composition table) of the specific beverage. For coffee and tea, the amount of cups per day was multiplied by number of teaspoons of sugar (6 g sugar per teaspoon) to each cup. Subsequently, all sugar content per day variables were summarised to total sugar from beverages per day.

Six variables with a very low variance (>90% of participants gave the same response) were excluded from the cluster analysis because of their tendency to form small and special clusters. The six variables were: coffee made with a cafetière; Greek or Turkish yoghurt; low-fat yoghurt; gratinated potatoes; cereals high in sugar; and dried fruits. Variables with a low variance (80–90% of participants gave the same response) were merged with similar foods if appropriate (11 variables were merged to five new variables; for complete overview see Appen ). Variables on portion size for beverages were excluded. Altogether, 55 variables on frequency of intake were included in the cluster analysis (Appen ). Clusters were extracted by Ward's method using squared Euclidian distance, in Predictive Analytics SoftWare (PASW) version 18 (SPSS Inc., Chicago, IL, USA). The values were not standardised as the distances between values were similar in all of the included variables. The number of clusters to derive was defined by the dendrogram and by controlling for robustness of different numbers of clusters through split-half replication. Cluster solutions ranging from 2 to 8 clusters were examined. Extraction of two or four clusters was interpreted as the best solution based on the dendrogram, were recognised through split-half replication and considered robust (32). When examining intake frequency of food items in two vs. four clusters, a cluster solution of four was considered appropriate as it produced more distinct dietary patterns (32, 33).

Biological markers and anthropometric variables were collected at gestational week 28±2 (visit 2). The methods for measurement of blood levels of glucose and glycosylated haemoglobin (HbA_1c) (23), fasting insulin, and C-peptide (34) have been described in detail elsewhere. Homeostasis model assessment of insulin resistance (HOMA-IR) were estimated by the Oxford University HOMA Calculator 2.2 with fasting glucose and C-peptide concentrations (35). Total cholesterol, HDL-cholesterol, LDL-cholesterol, and triacylglycerol (TAG) was measured using slide-adapted colorimetric method, observed at 540 nm (Vitros 5.1 FS, Ortho Clinical Diagnostics) at the Department of Multidisciplinary Laboratory Medicine and Medical Biochemistry, Akershus University Hospital.

Stature was measured to the nearest 0.1 cm using a fixed stadiometer (checked against a standard meter before the start of the study and twice yearly), and body weight and percent total body fat were measured with Tanita-BC 418 MA body composition analyser (22).

A linear stepwise (bidirectional) regression analysis was conducted to assess which food items that explained most of the variation in the clusters. Associations between intake frequencies of food items within the dietary patterns were analysed through Chi-square tests. Characteristics of the sample across dietary patterns were analysed with ANOVA if the variable was continuous and normally distributed, Kruskal-Wallis if continuous and non-parametric, and Chi-square tests if categorical. Multinomial regression analysis using main effects was performed to explore the association between ethnic origin and the dietary patterns, and adjusting for socio-economic and integration level. Fasting insulin and HOMA-IR were log-transformed to attain normally distributed data. Univariate general linear models (GLM) were executed to analyse the difference of biomarker levels within the dietary patterns while adjusting for age, percent total body fat, ethnic origin, socio-economic score, and integration score. PASW Statistics 18 (SPSS Inc., Chigago, IL, USA) was used in all statistical analyses.

Four robust dietary clusters were detected. Table 1 presents fractions of participants within each cluster with a frequency of intake above a given cut-off. Ten food items were excluded from this table because of either a low frequency of intake overall (fried fish, fish products, sugar-reduced jam, fish spreads for bread, natural and sweetened yoghurt) or low variance and insignificant differences between the dietary patterns (boiled potatoes, fruit juice, cutlets, and other fatty meats). The linear regression analysis showed that frequency of intake of 17 food items (of the 55 included in the cluster analysis) explained 58.7% of the variance within the dietary patterns, with the item full-fat milk (R ²=0.334) as the largest contributor. The other items were, in descending explanatory order: low-fat liver pâté and ham; skimmed milk; tea; coffee; beans and lentils; low-fat processed meat; semi-skimmed milk; egg as a sandwich spread; full-fat cheese; artificially sweetened soft drinks; lean fish; salty snacks; sweet biscuits; white bread; soft drinks with sugar; and chocolate.

Cluster 1 was characterised by frequent intake of full-fat milk, high sugar intake from beverages and frequent intake of dried fruits and nuts. The frequencies of eating fruit and vegetables were about average compared with the other dietary patterns, while these women had the most frequent intake of beans and lentils. The use of white bread and wholemeal bread in cluster 1 was equally frequent. Cluster 1 was also characterised by the most frequent intake of eggs as a sandwich spread, jam and chocolate spreads, sweet biscuits and sweet bakery products. Women in cluster 2 reported the most frequent intake of vegetables, the second most frequent intake of fruit and berries, and used mainly semi-skimmed milk. Daily use of bread was reported less frequent than that in cluster 4, but wholemeal bread was preferred. Cluster 2 was also characterised by frequent use of added sugar to tea or coffee, but a lower intake of confectionery and snacks as compared to the other three clusters. Women in cluster 3 reported the least frequent intake of fruits and vegetables and of any type of milk. Daily use of bread was also reported by the lowest proportion among women in cluster 3 and if used, white bread was preferred. Also, the frequency of using jam was high. Furthermore, the women in cluster 3 reported the most frequent intake of polished rice and pasta. The highest proportion daily eaters of bread was found in cluster 4 and the most frequent use of wholemeal bread. Subsequently, they also reported the most frequent use of spreads in general, but cheeses and meats were preferred over sweet spreads. Cluster 4 had the most frequent intake of semi-skimmed and skimmed milk, soft drinks with artificial sweeteners and coffee. Women in this cluster also reported the lowest intake of added sugars to tea and coffee, and had the most frequent intake of fruit and a relatively frequent intake of vegetables. However, these women also reported the most frequent intake of chocolate. Cluster 1 may be regarded the unhealthier dietary pattern of these four, followed by cluster 3. Cluster 4 could be considered relatively healthy overall, followed by cluster 2. However, both these dietary patterns had elements that could be considered unhealthy.

Selected characteristics of the sample grouped by dietary patterns are presented in Table 2. Small, but significant, differences in age and percent total body fat were found across the dietary patterns. Among first-generation immigrants only, the duration of residence was shortest in cluster 1, intermediate in clusters 2 and 3, and longest in cluster 4. The socio-economic score and integration score followed a similar pattern with lowest scores in cluster 1, but the large variation in SDs indicates diverse individual scores. Also, an association between ethnic origin and the dietary patterns was observed. European origin was mainly associated with cluster 4 (57.1%, n=201) and these women were least represented in cluster 1 (6%, n=21). Women of other ethnic origins fell almost equally into all four clusters, although origin from South and East Asia (11.7%, n=27) and the Middle East and Africa (12.6%, n=22) was least frequently observed in cluster 4.

Table 3 shows ORs of belonging to the dietary patterns by ethnic origin adjusted for age and percent body fat (Model 1), when additionally adjusting for socio-economic score (Model 2) and when adjusting for age, percent body fat, socio-economic score, and integration score (Model 3). Non-Europeans as compared to Europeans had higher ORs for belonging to clusters 1–3 vs. cluster 4. Women originating from the Middle East and Africa had an OR of 21.5 (95% CI 10.6–43.7), while South or East Asia had an OR of 21.1 (95% CI 11.0–40.4) of falling into the unhealthier cluster 1 vs. the healthier cluster 4, as compared to Europeans. In the fully adjusted model, the OR for the Middle East decreased to 3.8 (95% CI 1.6–8.9) and for South and East Asia to 6.5 (95% CI 3.1–13.5). This implies that a low socio-economic status and low integration score explained a large proportion of the observed ethnic differences. The fully adjusted ORs for being in clusters 1 and 2 vs. cluster 4 for women from The Middle East and Africa were considerably lower than women of South and East Asian women, but still significant. The pseudo R2 improved by each model (Table 3). The likelihood-ratio test for each of the models was significant (p<0.0001 in each model).

The relative healthiness of the dietary patterns was to some extent reflected in the biological markers before and after adjustment for ethnic origin (Table 4). Fasting glucose, 2-h glucose, fasting insulin, HbA_1c, and HOMA-IR were significantly different across the clusters before adjusting for ethnic origin, with the most beneficial values found in cluster 4 and no apparent trend between clusters 1, 2, and 3. However, total cholesterol varied significantly with the lowest levels found in clusters 2 and 3, while clusters 2 and 4 had significantly lower values for TAG. HDL- and LDL-cholesterol did not vary significantly across the patterns. After adjustment for ethnic origin, despite low power, fasting insulin and HOMA-IR remained significant markers, while TAG was borderline significant. When adjusting for socio-economic score and integration score instead of ethnic origin, to increase power, differences in fasting insulin, HbA_1c, and HOMA-IR remained significantly different across the dietary patterns, while fasting glucose and TAG was borderline significant (Table 5).