Diet during pregnancy: Influence of social characteristics and migration in the ELFE cohort

Abstract Better adherence to dietary guidelines during pregnancy is supposed to result in healthier perinatal outcomes. We aim to characterize the diets of pregnant women by hypothesis‐driven and exploratory approaches and describe potential social determinants. Analyses included 12 048 mothers from the French nationwide ELFE birth cohort. Dietary intake over the last three months of the pregnancy was assessed by a food frequency questionnaire. Two hypothesis‐driven scores (the Diet Quality score, based on benchmarks derived from the National Health and Nutrition Program Guidelines, and the PANDiet score, based on nutrient intake) were calculated. Exploratory dietary patterns were also identified by principal component analysis. Multiple linear regressions were used to assess associations of maternal social characteristics with dietary patterns, accounting for the possible effect modification by their migration status. Five dietary patterns were identified: the Western, Balanced, Bread and toppings, Processed products, and Milk and breakfast cereals. Younger maternal age, single motherhood, unemployment and the presence of older children in the household were related to a suboptimal diet during pregnancy. The less acculturated the women were, the healthier and less processed their diets were, independent of their socio‐economic position. Several social determinants of the quality of women's diets were however moderated by their migration status. These findings shed light on the relations between indicators of social vulnerability, such as single motherhood and unemployment, and poorer diet quality. Given the reduced diet quality that accompanies the acculturation process, it is of paramount importance to identify the specific factors or obstacles that affect migrant women in maintaining their diet quality advantage over the majority population.


| INTRODUCTION
Maternal diet is the major prenatal source of nutrients and can thus influence embryonic, placental, and fetal growth (Blumfield, Hure, Macdonald-Wicks, Smith, & Collins, 2013). Accordingly, maternal diet during pregnancy appears to be crucial for developing fetuses and their later-life health status (Abu-Saad & Fraser, 2010;McMillen et al., 2008). In particular, depletion of the fetal environment during pregnancy can induce a fetal adaptive response of the body, that prioritizes the growth of vital organs such as the brain over the development of others and leads to metabolic alterations (Hales & Barker, 2001). These adaptations are assumed to be especially harmful when followed by overnutrition of the newborn in the postnatal period. Beyond insufficient or excessive energy intake (Blumfield, Hure, Macdonald-Wicks, Smith, & Collins, 2012), the current problem in high-income countries is that of unbalanced diet and suboptimal micronutrient intake (Blumfield et al., 2013), especially in disadvantaged populations (Bertin et al., 2016;Lioret et al., 2012;Roberts, Cavill, Hancock, & Rutter, 2013).
Among the latter, cultural and ethnic minority groups, defined as migrants or population subgroups of migrant or indigenous background (Osei-Kwasi et al., 2016), are reported to have a higher prevalence of diet-related noncommunicable diseases, along with a suboptimal diet, compared with the native-born population (Faskunger, Eriksson, Johansson, Sundquist, & Sundquist, 2009;Patel et al., 2006). Beyond socio-economic position, various sociocultural factors have been shown to have a greater influence on dietary behaviours in cultural minority groups, including acculturation, a multidimensional process in which individuals adopt characteristic ways of life from the culture they are now living in that differ from the primary learning of their cultural background (D'Souza, Jayaweera, & Pickett, 2016;Osei-Kwasi et al., 2016). In their reviews, D'Souza and colleagues (D'Souza et al., 2016) noted that traditional pregnancy diets vary from society to society, and Urquia and colleagues (Urquia et al., 2010) noted that the association between foreign-born status and birth outcomes also varies and depends on both migrant subgroup and destination country.

Initiated in 2001, the French National Nutrition and Health
Program (PNNS) aims to improve the health status of the population by acting on one of its major determinants: nutrition (S. Hercberg, Chat-Yung, & Chauliac, 2008). This programme includes guidelines on dietary intake and physical activity for the general population, along with guidelines for specific populations, such as pregnant women.
Although several national surveys have monitored dietary intake and adherence to dietary guidelines in the French adult population (Bertin et al., 2016;Castetbon et al., 2009;Dubuisson et al., 2010;Volatier, 2000), these studies included too few pregnant women to assess their adherence to these recommendations. Assessing dietary patterns of pregnant women with specifically designed studies is an important public health objective, to make it possible to identify the groups at greatest risk so that they can receive more tailored information and interventions.
Two main classes of methods are used to derive dietary patterns: first, the production of diet quality scores or indices based on a priori benchmarks from nutritional guidelines or on nutrient intake, and second, the use of exploratory statistical methods, such as principal component analysis and cluster analysis (Kant, 2004). The latter a posteriori data reduction techniques are commonly used in nutritional epidemiology to summarize the variation in food intakes into a small number of patterns and have shown good reproducibility across studies (Newby & Tucker, 2004).
The objective of this study was to characterize the diet of pregnant women living in France, by using both hypothesis-driven (a priori) and exploratory or data-driven (a posteriori) approaches, and then to assess its association with demographic and socio-economic characteristics, while its moderation by migration status is considered.

| Study population
This analysis is based on data from the ELFE (Etude Longitudinale Française depuis l'Enfance) study, a multidisciplinary nationwide birth cohort including 18 329 children born in 2011 in a random sample of 349 maternity units in mainland France (Charles et al., 2019). Enrolment began in April 2011 and took place during 25 selected days, over four waves of 4-8 days each, each wave covering a different season. Infants met the inclusion criteria if they were singletons or twins born after 33 weeks of gestation to mothers aged 18 years or older who were not planning to move outside of metropolitan France in the next 3 years.
Overall, 51% of eligible parents agreed to their child's participation. Data were collected in standardized interviews conducted by trained interviewers and by self-administered questionnaires. Efforts

Key messages
• Beyond the well-known association of maternal education level and age with diet quality, these findings underline the relations of some indicators of social vulnerability, such as single motherhood and unemployment, with poorer diet quality during pregnancy.
• The less acculturated pregnant women were, the healthier and less processed their diets were, regardless of their socioeconomic position.
• Several demographic and socioeconomic determinants of the quality of women's diets were moderated by their migration status.
were made at inclusion to reach women who were not totally fluent in French: in particular, the information letter and consent form were also edited in the three languages most often spoken in France (after French), i.e. Arabic, Turkish and English. The interviews were also conducted in Soninké, Bambara and Wolof.
2.2 | Maternal diet during pregnancy 2.2.1 | Data collection At the hospital before discharge, the mothers enrolled in the study completed a self-administered food frequency questionnaire (FFQ) that described their dietary intake over the last three months of the pregnancy. Consumption of 125 food items was collected, including 12 nonalcoholic and 4 alcoholic beverages, with frequency assessed by a 7-item scale from 'never' to 'more than once a day'. Usual portion sizes for different food types were estimated with photos indicating five possible sizes, derived from the SU.VI.MAX portion book (S. Hercberg, Deheeger, & Preziosi, 2002) for 75 commonly eaten food items; and for the 50 remaining food items, the midportion in the SU.VI.MAX portion book was automatically assigned. After the transformation of FFQ frequency categories into daily frequencies, daily intake was assessed, for each food item, by combining frequency of intake and portion size. Nutrient intakes were then calculated for all women by multiplying the daily intake of each food item by the nutritional values in the SU.VI.MAX nutrient composition database (S. Hercberg, 2006). The 125 food items were then classified into 48 food groups. Women for whom more than 10 FFQ items were missing were excluded. Conversely, when 10 or fewer items were missing, the missing values were imputed by the sample median. Women were deemed likely misreporters of their food intake, and thus excluded, when their estimated total energy intake was less than 933 kcal/day (3rd percentile) or exceeded 5072 kcal/day (97th percentile). Women were also asked to report their frequency of intake of industrially processed foods on a five-item scale ranging from 'never' to 'always': reduced-sugar and reduced-fat products, grouped together in the diet food group; prepacked foods; canned foods; and ready-prepared dishes.
The pregnant women's adherence to the French nutritional guidelines (ANSES, 2016) was assessed with two complementary a priori approaches. The first, the Diet Quality score, was based on consumption of the main food groups, and the second, the PANDiet score, on a range of nutrient intakes. Calculation of the Diet Quality score therefore used 17 quantitative benchmarks (Table S1) adapted from the PNNS-GS (National Health and Nutrition Program Guideline Score) (Estaquio et al., 2009). The methodology applied the principle of the percentage of each guideline followed by women, consistent with the approach developed in the MoBa cohort (von Ruesten et al., 2014). The overall score, which can range from 0 to 17, is the sum of the scores obtained on each item: the higher the score, the better the adherence to the national guidelines. Details of this score have previously been published (Kadawathagedara et al., 2017). At the same time, the PANDiet score, previously developed and validated among French pregnant women (Bianchi, Mariotti, Verger, & Huneau, 2016), was calculated. It aims to measure the overall quality of a woman's diet by combining the probability of adequate intake for different nutrients, based on recommendations for women in the third trimester of pregnancy. The PANDiet score ranges from 0 to 100, with a higher score again indicating better nutritional quality.

| FFQ validation study
A separate study sought to validate the FFQ. It enrolled participants from maternity clinics during their appointment with their obstetrician in major cities across France. Of the 128 women informed about the study by their physician and asked to participate, 105 consented. Among them, 67 (age: 31.4 ± 4.8 years) could give the study dietician information for at least one 24-h recall over the telephone (out of three initially planned within the last three months of pregnancy); and 62 women completed the FFQ assessing foods consumed during the last trimester. Participants received the study FFQ by mail a month before the planned due date and a telephone reminder to complete the FFQ within a week after their baby's birth. Participants with an incomplete FFQ (n = 4) and those for whom the estimated energy intake by FFQ was considered unrealistic (n = 2; i.e. 600 Kcal for one subject and 6000 Kcal for the other subject) were excluded from the analysis. The analysis covered the data from the 56 women who provided at least one 24-h recall and the FFQ were analysed. The methodology of this validation study is described in greater detail in the Supporting Information, as are its results (Table S2).

| Demographic and socio-economic characteristics
As the demographic and socio-economic characteristics collected at the two-month telephone interview were more detailed than at the maternity ward interview, we prioritized the data collected at 2 months when they conflicted. Demographic and socio-economic characteristics studied were as follows: mother's age at delivery (18-25, 25-29, 30-34

| Sample selection
The study excluded infants whose parents withdrew consent during the first year (n = 57) ( Figure S1). In twin births, random selection of one twin (n = 287 exclusions) made it possible to count each mother of twins only once in the analysis. The analyses excluded women who did not complete the FFQ (n = 2,013), or with more than 10 missing values (n = 658), with implausible energy intake (n = 910) or with missing data for use of industrially processed products during pregnancy (n = 433); this left 13 881 women for dietary pattern identification.
The complete case analysis, which excluded women with at least one missing value for demographic and socioeconomic characteristics, had a sample of 12 048 women.

| Exploratory dietary patterns
Exploratory dietary patterns were derived by principal component analysis (PCA, FACTOR procedure; SAS software (Version 9.4, SAS Institute, Cary, NC, USA)) of the 46 standardized food groups consumed by more than half the sample. The number of factors to retain was determined with the diagram of eigenvalues (the Scree plot) and based on the interpretability of the factors (Cattell, 1966;Kline, 1994). A standardized score was calculated for each individual and pattern to quantify the individual's adherence to the pattern.
The association between each of the dietary scores, whether derived by hypothesis-driven or exploratory approaches, were analysed with Pearson's correlations. Recommendations for interpreting these correlation coefficients are: low 0.1-0.3, moderate 0.3-0.5 and high >0.5 (Cohen, 1988).

| Associations with demographic and socioeconomic characteristics
Associations of demographic and socio-economic characteristics with maternal hypothesis-driven and exploratory dietary patterns (outcomes) were tested by multiple linear regressions, which included all the demographic and socio-economic characteristics and were adjusted for variables related to the study design (maternity unit size and season of enrolment). Possible interactions (i.e. effect modifications) between migration status and all demographic and socioeconomic characteristics were tested for their relations with the various dietary patterns.

| Sensitivity analyses
To take into account the inclusion procedure and biases related to nonconsent, the data were weighted (Juillard, 2015) in sensitivity analyses. Weighting also included calibration on margins from the state register's statistical data and the 2010 French National Perinatal study (Blondel, Lelong, Kermarrec, Goffinet, & National Coordination Group of the National Perinatal Surveys, 2012) on the following variables: age, region, marital status, migration status, level of education and primiparity. This weighting was calculated for the subsample that completed the FFQ.
In addition to the main analyses of the complete cases, analyses with multiple imputations dealt with missing data for demographic and socio-economic characteristics (n = 13 618). Missingness of each variable was less than 11% (10.6% for household income; 9.5% for maternal education level; 6.9% for maternal migration status; 6.9% for maternal employment; 6.4% for single parenthood; 6.4% for parity; and 0% for all other variables of interest). On the assumption that data were missing at random, five independent datasets were generated with the fully conditional specification    Five dietary patterns were identified that together explained 28% of the total variance. The first pattern was characterized by high positive factor loadings for French fries, fast foods, potatoes, mixed dishes, cakes and pastries, sauce (e.g. mayonnaise or ketchup), sweet snacks, red meat, potato chips and similar products, processed meat, candies and breaded fish ( Table 2). It is labelled the 'Western' pattern.

| Ethical considerations
The second pattern was characterized by high positive factor loadings for cooked vegetables, raw vegetables, fish, fruit, whole-grain bread, legumes, yogurt and cottage cheese, and dried fruit and negative coefficients for sweetened beverages. It is referred to as the 'Balanced' pattern. The third pattern, with high coefficients for honey/jam, chocolate, butter and, to a lesser extent, cheese and bread, is named the 'Bread and toppings' pattern. The fourth pattern is labelled the 'Processed products' pattern in view of its high coefficients for prepacked foods, ready-prepared dishes, canned foods and diet foods.
Finally, the fifth pattern was characterized by high positive coefficients for milk, cocoa powder, and breakfast cereals and negative coefficients for coffee. It is called the 'Milk and breakfast cereals' pattern.
The correlation between the Diet Quality and the PANDiet scores was moderate (0.43, P < 0.0001). The correlations of both the Diet Quality and the PANDiet scores with the exploratory Balanced pattern score were respectively high (0.65, P < 0.0001) and moderate (0.42, P < 0.0001). The correlations of the remaining four exploratory dietary patterns with each of the Diet quality and the PANDiet scores were low, i.e. respectively −0.03 and 0.20 for the Western pattern (P > 0.05 and P < 0.0001); 0.01 and −0.05 for the Bread and toppings pattern (P > 0.05 and P < 0.0001); −0.03 and −0.14 for the Processed products pattern (both P < 0.0001); and 0.21 and 0.24 for the Milk and breakfast cereals pattern (both P < 0.0001).
3.2 | Demographic or socio-economic characteristics and maternal diet Table S3 presents the bivariate associations of the maternal dietary patterns with demographic or socio-economic characteristics.

Moderating effects of maternal migration status
There was no interaction between migration status and any of  (Table 3). Maternal occupation and single motherhood were not associated with it. Migrant women and, to a lesser extent, descendants of migrants, had lower scores on the Bread and topping pattern than the majority population did.

Milk and breakfast cereals dietary pattern
Maternal age and older children in the household were negatively related to this dietary pattern (   Values are estimates [95% CI] from multivariable linear regressions, also adjusted for region of residence, maternity unit size and enrolment wave; the global significance was indicated with the following labels: *** for p < 0.001, ** for p < 0.01 and * for p < 0.05. b Corresponding to ≤$815/month.
F I G U R E 1 Multivariate associations between demographic or socio-economic characteristics and maternal diet by maternal migration status: The ELFE study, 2011. Figure 1a presents the multivariate associations among migrants (n = 853), Figure 1b those among descendants of migrants (n = 1195), and Figure 1c those among the majority population (n = 10 000). All multivariable linear regressions presented in this figure were adjusted for region of residence, maternity unit size and recruitment wave.

Processed dietary pattern
None of maternal age, maternal education level and maternal occupation was associated with this dietary pattern (Figure 1). Household income was positively associated with it in the majority population and in descendants of migrants, but not in migrants. The presence of older children was associated with greater adherence to this pattern in descendants of migrants only. Single motherhood was associated with a higher score on this pattern, regardless of migration status.
The analyses based on weighted data or on multiple imputations of missing data produced very similar findings (data not shown).

| DISCUSSION
This study is the first to provide insights into the demographic and socio-economic determinants of pregnant women's diet in France, while taking the possible moderating effects of migration status into account. Five distinct dietary patterns were identified and labelled as Western, Balanced, Bread and toppings, Processed products, and Milk and breakfast cereals. Overall, older age and foreign cultural background were related to better quality diets during pregnancy, whether quality was assessed by hypothesis-driven or exploratory approaches.
The reverse was true for single motherhood, unemployment, and older children in the household, all of which were associated with suboptimal diets. Maternal education and household income were negatively associated with the Western dietary pattern in the overall population, while their association with the healthier dietary patterns was moderated by the mother's migration status: their traditional positive association with a healthy diet was essentially confirmed in the majority population, and to a lesser extent in mothers from different cultural backgrounds. This acculturation process was also suggested for the Processed dietary pattern, which had higher scores in households with higher incomes, in the majority population and to a lesser extent in descendants of migrants, but not in migrant mothers.
Comparison of the maternal and household characteristics associated with maternal diet in a US study (Bodnar & Siega-Riz, 2002) showed that a high score on the Diet Quality Index for Pregnancy was positively associated with women's age, education, and income. Similar relations between women's characteristics and diet quality during pregnancy have been reported in Norway (Hillesund, Bere, Haugen, & Overby, 2014) and Spain (Rodriguez-Bernal et al., 2010). The Norwegian study demonstrated that adherence to Nordic or Norwe- associations were also observed with unemployment and older children in the household. In a context of higher social vulnerability, obstacles to the availability and accessibility of healthy food, along with time constraints, may partly explain these associations (Dlugonski & Motl, 2013;Mills et al., 2017;Pollard, Kirk, & Cade, 2002) and should inform nutritional interventions targeted at more at-risk families.
The Diet Quality score was constructed to consider adherence to guidelines for all of the core food groups. These groups, however, Only few studies have combined hypothesis-driven and exploratory approaches to characterize maternal diet. Principal component analysis is a standard technique used in population studies to explore dietary patterns (Bertin et al., 2016;Lioret et al., 2012;Northstone, 2012;Northstone, Emmett, & Rogers, 2008), which takes the potential effect of interactions within combinations of foods and nutrients into account (Newby & Tucker, 2004). While the hypothesis-driven approach promotes international comparisons and enables the exploration of global adherence to nutritional guidelines, the exploratory approach, which identifies specific patterns within a country and various dimensions of diet quality, may enable the provision of more tailored food-based recommendations.
It should be noted that the two complementary hypothesis-driven quality scores used in the current study (one derived from food group consumption and the other one based on nutrient intake) were moderately correlated (r = 0.43). Both scores were also strongly correlated with the Balanced dietary pattern. These correlations support the external validity of the hypothesis-driven and exploratory scores.
The ELFE-FFQ provided reasonably good estimates of food and nutrient intake compared with repeated 24-h recalls (Table S2); correlation coefficients are similar to those reported in other validation studies among pregnant women (Cheng et al., 2008;Erkkola et al., 2001;Mikkelsen, Osler, & Olsen, 2006;Mouratidou, Ford, & Fraser, 2006;Pinto et al., 2010;Robinson, Godfrey, Osmond, Cox, & Barker, 1996), although a US study noted higher correlations (Brown et al., 1996). There is no general consensus on a 'satisfactory' level of correlation. In most validation studies, correlation coefficients between methods are considered poor when <0.30; fair when between 0.3 to 0.49; and good when >0.50 (Hankin, Wilkens, Kolonel, & Yoshizawa, 1991). Adjustment for energy and correction for random intraindividual variation tended to increase the magnitude of the correlation coefficients (Brantsaeter, Haugen, Alexander, & Meltzer, 2008;Brown et al., 1996;Erkkola et al., 2001;Mouratidou et al., 2006). Correlation coefficients are commonly employed and reported in validation studies, but the level of agreement assessed by cross-classification into quartiles or quintiles of their intake is considered more appropriate to evaluate the ability of an FFQ to rank individuals into low or high intakes relative to the reference method (Garrow, 1995). In the ELFE study, only a very small percentage of the women was grossly misclassified (classified into opposing quintiles in FFQ vs. 24-h recalls). It should be noted that a review of 90 studies of energy intake during pregnancy in high-income countries found an average intake of 2340 Kcal in the third trimester, very close to the FFQ estimates reported here (Blumfield et al., 2012). The mean PANDiet score was also very close to the value previously estimated in a population of pregnant women or in any case of childbearing age (Bianchi et al., 2016). It is however recognized that the FFQ was completed after delivery and collected dietary intake retrospectively over the last three months of pregnancy, which does not necessarily reflect the diet across all the pregnancy period.
The ELFE cohort is a nationwide study of births in 2011 in metropolitan France (excluding very premature babies). The main analyses were conducted on the complete case sample. It must also be acknowledged that, as it is often the case in cohorts, women not participating in the cohort or those excluded from the analyses were more strongly socially disadvantaged and more likely to have been born abroad, which is a potential limitation in terms of generalizability of the findings. We also acknowledge that whereas the main documents were edited in the three languages most often spoken in

| CONCLUSION
The present study used both hypothesis-driven and exploratory approaches to characterize the diet of pregnant women more comprehensively. In addition to two scores indicating higher adherence to dietary (Diet quality score) and nutritional (PANDiet score) guidelines, five distinct data-driven dietary patterns were identified and labelled as Western, Balanced, Bread and toppings, Processed products, and Milk and breakfast cereals. These findings from the large, nationwide ELFE birth cohort suggest that younger maternal age, single mother- Research Agency under the "Investissements d'avenir" programme (ANR-11-EQPX-0038). The funders had no role in the study design, data collection and analysis, decision to publish or preparation of the manuscript.

DATA AVAILABILITY STATEMENT
The data underlying the findings cannot be made freely available for ethical and legal restrictions imposed, because this study includes a substantial number of variables that, together, could be used to reidentify the participants based on a few key characteristics and then be used to have access to other personal data. Therefore, the French ethics authority strictly forbids making these data freely available.
However, they can be obtained upon request from the ELFE principal investigator. Readers may contact marie-aline.charles@inserm.fr to request the data.

CONFLICTS OF INTEREST
The other authors had no conflicts of interest relevant to this article to disclose.

CONTRIBUTIONS
MK conceptualized and designed the study, conducted statistical analyses, drafted the initial manuscript, and approved the final manuscript as submitted; MND and AF designed the data collection instruments, supervised data collection, critically reviewed the manuscript, and approved the final manuscript as submitted; NA, MAC, SL and BLG conceptualized and designed the study, contributed to data analysis and interpretation of the study, reviewed and revised the manuscript, and approved the final manuscript as submitted.