Diet during pregnancy is crucial for maternal and child health( Reference Barker, Gelow and Thornburg 1 , Reference Ramakrishnan, Grant and Goldenberg 2 ). Energy and nutrient intakes must support growth of maternal and fetal tissues and accumulate reserves for lactation( Reference Douglas 3 ). Inadequate nutrition (deficits and excesses) bears the risk of permanent consequences for the offspring( Reference Robinson 4 ). Life course epidemiology frames pregnancy as a critical period( 5 ). Pregnancy has been identified as a period with great potential for change in dietary habits( Reference Hoffmann, Nunes and Schmidt 6 ). Heightened awareness of potential threats to own and child’s health may motivate women to adapt health-promoting behaviours including nutritional changes( Reference Gardner, Croker and Barr 7 ).
Dietary assessment is complex; it involves recording and analysing a multitude of foods and drinks consumed every day and in varying quantities( Reference Trichopoulos and Lagiou 8 ). Due to this complexity, diet is methodologically difficult to capture and no gold-standard method exists to date( Reference Trautwein and Hermann 9 ). Growing concerns about the limitations of examining single foods or nutrients in isolation( Reference Hu, Rimm and Smith-Warner 10 ) led to the development of the concept of dietary patterns (DP) 30 years ago( Reference Slattery 11 ).
Dietary quality (DQ) is another relatively new concept to capture diet as a whole by scoring adherence to (national) dietary guidelines, rating the diversity of food choice in key food groups, or scoring predefined food patterns known to protect or impair health( Reference Wirt and Collins 12 ).
Their definitions overlap. DP have been defined as ‘the quantities, proportions, variety, or combination of different foods, drinks, and nutrients (when available) in diets, and the frequency with which they are habitually consumed’ (p. 1)( Reference Essery Stoody, Spahn and McGrane 13 ). Similarly, DQ has been described as a ‘relatively new concept [that] involves the assessment of both quality and variety of the entire diet, enabling examination of associations between whole foods and health status, rather than just nutrients’ (p. 2473)( Reference Wirt and Collins 12 ).
A review of the health effects of gestational DP identified a range of health outcomes of mothers (e.g. infertility, gestational diabetes mellitus and depressive symptoms) and their children (e.g. fetal growth, preterm birth and risk of asthma)( Reference Chen, Zhao and Mao 14 ). Likewise DQ was associated with blood TAG( Reference Martin, Siega-Riz and Sotres-Alvarez 15 ), pre-eclampsia( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 ) and fetal growth restriction( Reference Rodriguez-Bernal, Rebagliato and Iniguez 17 ).
Considering the importance of diet in pregnancy and the increase in studies assessing diet as patterns or quality, a systematic review of its determinants is necessary in order to assess the population needs and develop effective public health interventions.
Methods
Research question and concepts
We previously developed a conceptual framework of determinants of diet in pregnancy taking a multiple-determinants and life course view (described below). We conducted the present systematic review to test the ‘fit’ of our framework and summarise the available evidence.
Diet in this context was defined as a representation of overall diet, not merely single foods or nutrients, using DP, DQ or guideline compliance. We recognise that these are distinct entities, but they have the same underlying principle: to capture total dietary intake as best as possible.
We understood determinant according to Last’s definition of ‘any factor, whether event, characteristic, or other definable entity, that brings about change in a health condition, or other defined characteristic’ (p. 37)( Reference Last 18 ). In our case we considered factors which brought about change in diet (measured as DP, DQ or guideline adherence) of pregnant women.
Our conceptual framework (Fig. 1) is based on an initial literature scoping, the Conceptual Framework for Patterns of Determinants of Health( Reference Evans and Stoddart 19 ) and the Perinatal Health Framework( Reference Misra 20 ) (which built on the former but added the angle of time and adapted it for the case of perinatal health). The framework includes different determinants: environmental, sociodemographic and individual responses, these are not limited to the perinatal period, and pregnancy-related factors.
Conceptual multiple-determinants life course framework of diet in pregnancy (DP, dietary pattern; DQ, dietary quality)
Determinants are positioned based on their distance to diet. As distal we classed environmental determinants. We expanded the meaning of environment beyond the physical environment to include the categories set out in the environmental research framework for weight-gain prevention( Reference Sleddens, Kroeze and Kohl 21 ) as consisting of physical, sociocultural, economic/financial and political factors. We amended this categorisation: the categories of political and economic determinants were merged and the category of medical environment was added. We hypothesised that the medical/health-care system plays a greater role in pregnancy due to more frequent contact with practitioners. For example, health-care practitioners may influence diet through information giving. Conversely, inadequate dietary intake may result in contacts with health-care providers to get treatment or advice for nutrition-related health problems such as anaemia.
As proximal determinants we defined the remaining categories. Sociodemographic factors include age, education, employment, ethnicity and other personal attributes like partnership. They may impact diet directly or indirectly via effects on individual response or pregnancy.
Evans and Stoddard argue that the social and physical environment can lead to individual differences in biological response (e.g. expression of genes) or behavioural response (e.g. engaging in health-risk behaviours in response to stress)( Reference Evans and Stoddart 19 ). Misra also outlined that negative health behaviour may occur in response to experiences such as discrimination( Reference Misra 20 ). The review of individual responses is relevant for the study of dietary intake as well. Research indicates that genetic differences may explain diverse biological responses to overfeeding( Reference Bouchard, Tremblay and Després 22 ) and individuals differ in their behavioural response to internal and external food cues( Reference Schüz, Schüz and Ferguson 23 ). We hypothesised individual responses to be influenced by both individual and environmental factors and thus positioned this category in between both these categories, distance wise.
Finally, pregnancy factors were defined as those determinants that relate to pregnancy, or only act during pregnancy, or mediate, or modify the influence of existing determinants during pregnancy. Pregnancy is known to influence dietary intake due to physical symptoms such as nausea and food aversions as well as psychological factors such as higher or lower restraint of eating in anticipation of gestational weight gain( Reference Anderson 24 ). In distance terms pregnancy-related factors were placed closest to diet in pregnancy, since pregnancy is described as a unique period and was thus considered the most immediate influence.
Data sources and search strategy
Four databases were searched from the date of their inception to January 2015; searches were updated in April 2016 (see Appendix for search strategies). The search combined three concepts: ‘determinants’, ‘dietary patterns’ and ‘pregnancy’. Search terms were amended slightly for each database. In addition, hand searches were performed (reference lists of all obtained articles, table of contents of key journals and conference abstracts). One publication had to be excluded from the review as crucial data were missing in the original publication( Reference Fowles and Gabrielson 25 ) and could not be retrieved.
Study selection
Inclusion and exclusion criteria are assembled in Table 1. Briefly, studies had to be of observational design and published in English or German. Participants had to be from a high- or upper-middle-income country( 26 ), i.e. countries with generally an abundance of food, where dietary intake is a reflection of food choice or access rather than availability. Participants had to be pregnant or in postpartum at the time of dietary assessment and the measurements had to refer to any time during pregnancy. The study’s aim had to be the assessment of determinants of diet in pregnancy. Determinants of diet could be sociodemographic, individual responses, environmental or pregnancy-related factors. Dietary intake had to be reported as DP or DQ, which included measurements of adherence to dietary guidelines.
Inclusion and exclusion criteria
PKU, phenylketonuria.
Data extraction and analysis
Screening of articles and data extraction was conducted in two steps. First, relevant studies were identified based on title and abstract by one reviewer (I.-M.D.). All 130 articles that could not clearly be excluded beyond doubt were read in full by two reviewers (I.-M.D. and B.B. or A.G.), results were compared and disagreements resolved by discussion. A data extraction form was designed, piloted and adjusted. One reviewer (I.-M.D.) extracted data in consultation with the co-authors and a statistician if study reports were unclear.
Assessment of risk of bias
The Newcastle–Ottawa Scale (NOS) for assessing the quality of non-randomised studies in meta-analyses was used to assess the likelihood of bias in each publication included( Reference Wells, Shea and O’Connel 27 ). The NOS has been recommended for use in reviews of observational studies( Reference Lo, Mertz and Loeb 28 ). We adapted the NOS to fit the purpose of the current review (Table 2).
Adapted Newcastle–Ottawa Scale for assessing the quality of non-randomised studies
SES, socio-economic status.
† We awarded a * for ≥3d dietary records if participants were trained in record keeping/records were interviewer-checked, or any method described as validated by the authors.
We use the terms ‘risk of bias’ and ‘study quality’ interchangeably and not as a judgement of the authors’ methodological merits, but rather of how ‘relevant’ the study was for our review.
Statistical methods
The studies that were included were heterogeneous in sample size, population and methods used for assessing diet (Table 3). When studies show heterogeneity on so many levels a pooling of results (meta-analysis) is not appropriate but a narrative synthesis can be conducted, whereby studies are narratively described, trends explored and reasons for inconsistencies of findings discussed( Reference Denison, Dodds and Ntani 29 ).
Characteristics of studies included in the present review
S, selection; C, comparison; O, outcome (for cohort studies); E, exposure (for case–control studies); ALSPAC, Avon Longitudinal Study of Parents and Children; CANDLE, Conditions Affecting Neurocognitive Development and Learning in Early Childhood; DIPP, Type 1 Diabetes Prediction and Prevention Project; ECCAGE, Study of Food Intake and Eating Behavior during Pregnancy (Brazil); PHP, Prenatal Health Project; PIN, Pregnancy, Infection, and Nutrition Study; CH, cohort study; CS, cross-sectional study; DP, dietary pattern(s); GW, gestational week; CFG, Eating Well with Canada’s Food Guide; DQ, dietary quality; MD, Mediterranean diet; HEI, Healthy Eating Index; WIC, Special Supplemental Nutrition Program for Women, Infants, and Children; CFG adh., adherence to Eating Well with Canada’s Food Guide; MD adh., Mediterranean diet adherence (score).
* Study quality aspects: Newcastle–Ottawa-Scale rating (actual score/maximum score).
† Diet measured at several time points over the course of pregnancy (all other studies: only one single measurement).
Results
A total of 4368 articles were identified, 4238 articles were excluded based on their title and abstract (Fig. 2). Accordingly, 130 full-text publications were read of which seventeen met all inclusion criteria. They presented results of twelve studies. All were written in English. No abstracts of unpublished studies were identified. Nine studies were published in the past 5 years, indicating that this is a new research area.
Flowchart showing the selection of studies for the present review on determinants of dietary patterns and diet quality during pregnancy
Most studies were from North America (n 6) or Europe (n 5), one was from South America. Eight were cohort studies and four were cross-sectional studies. Sizes ranged from fifty to 12 053 participants (Table 3).
Five of the seventeen publications assessed diet using DP( Reference Hoffmann, Nunes and Schmidt 6 , Reference Northstone, Emmett and Rogers 30 – Reference Arkkola, Uusitalo and Kronberg-Kippila 33 ). A further publication assessed DP with adherence scores to the Mediterranean diet( Reference Kritsotakis, Chatzi and Vassilaki 34 ). DQ was assessed in ten of the seventeen publications( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 , Reference Nash, Gilliland and Evers 35 – Reference Watts, Rockett and Baer 43 ) using different DQ indices (Table 3); one publication assessed DQ using guideline adherence( Reference Fowler, Evers and Campbell 44 ).
Different DQ tools were used but in all higher scores indicated higher quality. As anticipated, the assessment of DP was more diverse. Some studies used adherence scores where higher scores indicated higher adherence, some classed participants into mutually exclusive DP groups.
The NOS scores for cohort studies ranged from 5 to 8 (maximum 9). For publications of the three cross-sectional studies, NOS scores ranged from 5 to 7 (maximum: 10; Table 3).
Determinants of diet in pregnancy among reviewed publications (n 17)
Table 4 shows the different factors assessed in each study. The sociodemographic factor most frequently investigated was education( Reference Hoffmann, Nunes and Schmidt 6 , Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 , Reference Northstone, Emmett and Rogers 30 – Reference Fowles, Bryant and Kim 39 , Reference Fowles, Stang and Bryant 41 , Reference Tsigga, Filis and Hatzopoulou 42 , Reference Fowler, Evers and Campbell 44 ), followed by age( Reference Hoffmann, Nunes and Schmidt 6 , Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 , Reference Northstone, Emmett and Rogers 30 – Reference Kritsotakis, Chatzi and Vassilaki 34 , Reference Bodnar and Siega-Riz 36 – Reference Fowles, Bryant and Kim 39 , Reference Tsigga, Filis and Hatzopoulou 42 , Reference Watts, Rockett and Baer 43 ).
Determinants of diet during pregnancy identified in the present review
ALSPAC, Avon Longitudinal Study of Parents and Children; CANDLE, Conditions Affecting Neurocognitive Development and Learning in Early Childhood; DIPP, Type 1 Diabetes Prediction and Prevention Project; ECCAGE, Study of Food Intake and Eating Behavior during Pregnancy (Brazil); PHP, Prenatal Health Project; PIN, Pregnancy, Infection, and Nutrition Study; GW, gestational week; TM, trimester; DP, dietary pattern; PCA, principal component analysis; EFA, exploratory factor analysis; CFG, Eating Well with Canada’s Food Guide; NCI, National Cancer Institute (USA); DQI-P, Diet Quality Index for Pregnancy; MD, Mediterranean diet; HEI, Healthy Eating Index; SCA, simple correspondence analysis; HSFFQ, Harvard Service Food Frequency Questionnaire; AHEI-P, Alternate Healthy Eating Index adapted for pregnancy; HH, household; mPAL, measured physical activity level.
Other commonly measured sociodemographic determinants were ethnicity/birthplace/nationality( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 , Reference Northstone, Emmett and Rogers 30 , Reference Völgyi, Carroll and Hare 31 , Reference Kritsotakis, Chatzi and Vassilaki 34 – Reference Laraia, Siega-Riz and Kaufman 38 , Reference Fowles, Stang and Bryant 41 , Reference Watts, Rockett and Baer 43 , Reference Fowler, Evers and Campbell 44 ), income/financial difficulty/Medicaid( Reference Northstone, Emmett and Rogers 30 , Reference Völgyi, Carroll and Hare 31 , Reference Bodnar and Siega-Riz 36 – Reference Laraia, Siega-Riz and Kaufman 38 , Reference Fowles, Stang and Bryant 41 – Reference Fowler, Evers and Campbell 44 ) and marital status/partnership/cohabitation( Reference Hoffmann, Nunes and Schmidt 6 , Reference Northstone, Emmett and Rogers 30 , Reference Völgyi, Carroll and Hare 31 , Reference Kritsotakis, Chatzi and Vassilaki 34 , Reference Nash, Gilliland and Evers 35 , Reference Laraia, Bodnar and Siega-Riz 37 , Reference Laraia, Siega-Riz and Kaufman 38 , Reference Fowles, Stang and Bryant 41 , Reference Fowler, Evers and Campbell 44 ). Occupation/employment( Reference Hoffmann, Nunes and Schmidt 6 , Reference Northstone, Emmett and Rogers 30 , Reference Fowler, Evers and Campbell 44 ) was less commonly assessed.
The most frequently used individual response factors were pre-pregnancy BMI or weight category( Reference Hoffmann, Nunes and Schmidt 6 , Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 , Reference Northstone, Emmett and Rogers 30 – Reference Cucó, Fernández-Ballart and Sala 32 , Reference Laraia, Bodnar and Siega-Riz 37 , Reference Fowles, Timmerman and Bryant 40 – Reference Tsigga, Filis and Hatzopoulou 42 , Reference Fowler, Evers and Campbell 44 ), smoking before( Reference Cucó, Fernández-Ballart and Sala 32 , Reference Fowles, Stang and Bryant 41 , Reference Watts, Rockett and Baer 43 ) and during pregnancy( Reference Northstone, Emmett and Rogers 30 , Reference Cucó, Fernández-Ballart and Sala 32 – Reference Kritsotakis, Chatzi and Vassilaki 34 , Reference Laraia, Bodnar and Siega-Riz 37 , Reference Fowler, Evers and Campbell 44 ), and physical activity/exercise before( Reference Kritsotakis, Chatzi and Vassilaki 34 ) and during pregnancy( Reference Northstone, Emmett and Rogers 30 , Reference Cucó, Fernández-Ballart and Sala 32 , Reference Nash, Gilliland and Evers 35 , Reference Laraia, Bodnar and Siega-Riz 37 , Reference Fowles, Stang and Bryant 41 , Reference Fowles, Stang and Bryant 41 , Reference Fowles, Stang and Bryant 41 ). Other aspects of health behaviour were less commonly assessed such as supplement use( Reference Laraia, Bodnar and Siega-Riz 37 , Reference Fowler, Evers and Campbell 44 ), alcohol during pregnancy( Reference Fowler, Evers and Campbell 44 ) and caffeine during pregnancy( Reference Fowler, Evers and Campbell 44 ).
The most often assessed pregnancy-related determinant was parity( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 , Reference Northstone, Emmett and Rogers 30 , Reference Völgyi, Carroll and Hare 31 , Reference Arkkola, Uusitalo and Kronberg-Kippila 33 – Reference Laraia, Bodnar and Siega-Riz 37 , Reference Tsigga, Filis and Hatzopoulou 42 , Reference Fowler, Evers and Campbell 44 ); studies also assessed the influence of nausea( Reference Nash, Gilliland and Evers 35 ) and pregnancy body image( Reference Northstone, Emmett and Rogers 30 ).
The category of environmental factors was considered in only a few publications, which mainly looked at the living environment/place of residence( Reference Arkkola, Uusitalo and Kronberg-Kippila 33 , Reference Kritsotakis, Chatzi and Vassilaki 34 , Reference Tsigga, Filis and Hatzopoulou 42 , Reference Watts, Rockett and Baer 43 ), social environment (support)( Reference Kritsotakis, Chatzi and Vassilaki 34 , Reference Nash, Gilliland and Evers 35 , Reference Fowles, Bryant and Kim 39 ) and food environment( Reference Nash, Gilliland and Evers 35 ).
The influence of depression( Reference Northstone, Emmett and Rogers 30 , Reference Fowles, Timmerman and Bryant 40 ) and stress/anxiety( Reference Northstone, Emmett and Rogers 30 , Reference Nash, Gilliland and Evers 35 , Reference Fowles, Bryant and Kim 39 , Reference Fowles, Timmerman and Bryant 40 ) also emerged as determinants. As these did not fit any of the four categories of determinants we grouped them into a new category, psychological health, which could be regarded as an individual psychological response.
Studies reviewed (n 12)
The ALSPAC (Avon Longitudinal Study of Parents and Children) cohort (UK) benefited from a large sample, the assessment of a multitude of determinants and the use of multivariable-adjusted analyses( Reference Northstone, Emmett and Rogers 30 ). DP were derived using principal component analysis, a type of factor analysis which aims to reduce food variables to underlying factors (DP) that explain as much variation in the data as possible( Reference Reedy, Wirfalt and Flood 45 ). All five patterns combined explained only 31·3 % of variability, which may be a reflection of the number of variables analysed or an indication that further unidentified latent DP exist in that population( Reference Hu, Rimm and Smith-Warner 10 ).
The CANDLE (Conditions Affecting Neurocognitive Development and Learning in Early Childhood) cohort included predominantly African-American women from the southern USA( Reference Völgyi, Carroll and Hare 31 ). A third of the sample was obese and a quarter overweight. The diet assessment covered a 3-month period making recall bias a possibility, although administering the FFQ by trained interviewers may have helped to overcome this issue. The associations between determinants and DP were not adjusted for potential confounders.
The study by Cucó et al. (Spain) was the only study with a longitudinal analysis( Reference Cucó, Fernández-Ballart and Sala 32 ). Diet was measured using 7d records assessed by trained interviewers; a method we considered would reduce bias from recall or under-reporting. The sample was small and consisted of women who were more educated than representative for that geographical area. The association between patterns and determinants was assessed by fitting multiple linear regression models. The explained variance for both patterns was low, at 11 to 15 % for ‘Sweetened beverages and sugars’ and 9 to 11 % for ‘Vegetables and meat’ across the different time points. This may be explained due to a large number of variables in relation to sample size or indicate the existence of further unidentified patterns( Reference Hu, Rimm and Smith-Warner 10 ). Exploratory factor analysis tends to work better with larger sample sizes( Reference Yong and Pearce 46 ) and more heterogeneous samples( Reference Fabrigar, Wegener and MacCallum 47 ). Given the effort that went into that study and its longitudinal nature, it is a shame (for the aim of our review) that only four determinants made it into the final adjusted model: smoking, physical activity, age and BMI.
The DIPP (Type 1 Diabetes Prediction and Prevention Project) cohort benefited from a validated FFQ that was adapted to fit the Finnish diet( Reference Arkkola, Uusitalo and Kronberg-Kippila 33 ). Women received the questionnaire after birth; this information was checked by an interviewer but only 3 months later. This could have led to recall bias. Non-response bias is also possible since women who did not complete dietary information had lower education but higher parity. Seven DP were identified through principal component analysis. Collectively these patterns explained 29·5 % of variance, therefore the variance explained by each individual pattern was low. Factor loadings of 0·2 or greater were considered in pattern derivation. This is lower than recommended( Reference Yong and Pearce 46 ) and may lead to DP that lack construct validity( Reference Castro, Baltar and Selem 48 ). A pattern such as ‘Alcohol and butter’ is not intuitively understandable, and the presence of alcohol in the diet of pregnant women is startling. Interpretation of the ‘Healthy’ dietary pattern is more straightforward. As the multiple linear regression analysis was not adjusted, it is difficult to exclude the presence of confounding.
The Brazilian ECCAGE (Study of Food Intake and Eating Behavior during Pregnancy) Study was the only publication using cluster analysis( Reference Hoffmann, Nunes and Schmidt 6 ). This approach differs from the commonly used factor analyses as participants are ‘clustered’ in accordance with similarities in their dietary intake rather than foods being ‘factored’ that correlate greatly( Reference Reedy, Wirfalt and Flood 45 ). The study satisfied all requirements regarding sample selection and representativeness, but analyses were not adjusted for confounders. The FFQ was validated for use in pregnancy but validity was found to be low; dietary intake may thus not have been adequately captured.
Two publications from the Canadian PHP (Perinatal Health Project) assessed a range of determinants. In the first publication only parity was associated with meeting guidelines. Given that only 3·5 % of participants were classed as guideline compliant, the ability to assess differences between the compliant and non-compliant may have become impaired through lack of power( Reference Fowler, Evers and Campbell 44 ). The second publication identified more factors that were associated with the Diet Quality Index adapted for pregnancy (DQI-P)( Reference Nash, Gilliland and Evers 35 ). Overall the model had a low R 2, indicating that only a small proportion of variability in DQ was explained by the measured determinants( Reference Nash, Gilliland and Evers 35 ).
The PIN (Pregnancy, Infection, and Nutrition) study was represented by three publications( Reference Bodnar and Siega-Riz 36 – Reference Laraia, Siega-Riz and Kaufman 38 ). The FFQ used in PIN has been shown to underestimate grain servings, which may have biased the results( Reference Laraia, Bodnar and Siega-Riz 37 ). Results on food environment are limited by the fact that distance to food retail is a rather crude measure of access( Reference Laraia, Siega-Riz and Kaufman 38 ) and factors besides access, such as income, may also influence food purchasing. In two publications analyses were adjusted for confounders( Reference Laraia, Bodnar and Siega-Riz 37 , Reference Laraia, Siega-Riz and Kaufman 38 ).
The US Project Viva study used an FFQ specifically validated for use in pregnant women and both crude and adjusted analyses to assess the associations between determinants and diet( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 ). However, generalisability from this cohort may be impaired due to higher than average socio-economic position and lower prevalence of overweight and obesity than the national average. Bias may have resulted from determinants being assessed by self-report rather than validated by interviewer assessment.
Women of Greek origin (rather than immigrants) and those with higher education were over-represented in the Rhea cohort, limiting generalisability of findings. The study benefited from established scales for assessing social capital and dietary intake, as well as analyses that were adjusted for a wide range of confounders( Reference Kritsotakis, Chatzi and Vassilaki 34 ).
Three publications of a US study of low-income women point to the involvement of stress, distress and anxiety on DQ. However, that study consisted of a convenience sample of women recruited through a small number of clinics offering free services to un- and underinsured pregnant women, deeming the sample not representative. Results are published on similar topics but corresponding to fifty( Reference Fowles, Timmerman and Bryant 40 ), seventy-one( Reference Fowles, Stang and Bryant 41 ) and 118( Reference Fowles, Bryant and Kim 39 ) participants; it thus seems like data were analysed before participant recruitment was completed, which could have biased later analyses. The study includes the only publication reporting a sample size calculation; sample size was adequate for the latest publication.
When interpreting results from the studies by Tsigga et al.( Reference Tsigga, Filis and Hatzopoulou 42 ) and Watts et al.( Reference Watts, Rockett and Baer 43 ), readers must be aware that both are cross-sectional studies with rather low NOS rating. Potential sources of bias include not reporting sample size calculation( Reference Tsigga, Filis and Hatzopoulou 42 ), not reporting participation rate and analyses not being adjusted or stratified( Reference Tsigga, Filis and Hatzopoulou 42 , Reference Watts, Rockett and Baer 43 ). Also, Tsigga et al. used the Healthy Eating Index (HEI) without adaptations for pregnancy and Watts et al. adapted the DQI-P; in neither case is it clear if these modifications of instruments (or lack thereof) are appropriate to capture diet in the target population.
Determinants reviewed in seventeen publications
Pregnancy-related
Parity was the most commonly investigated pregnancy-related factor (ten publications). In the ALSPAC and DIPP studies, pattern scores were associated with parity( Reference Northstone, Emmett and Rogers 30 , Reference Arkkola, Uusitalo and Kronberg-Kippila 33 ). In the PHP study, parity was associated with meeting guidelines( Reference Fowler, Evers and Campbell 44 ) and DQ score( Reference Nash, Gilliland and Evers 35 ). In the US cohorts PIN( Reference Bodnar and Siega-Riz 36 ) and Project Viva( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 ), parity was inversely associated with DQ. The same was observed for mean Mediterranean diet score in Rhea participants( Reference Kritsotakis, Chatzi and Vassilaki 34 ). Another Greek study found that parity did not appear to influence HEI score( Reference Tsigga, Filis and Hatzopoulou 42 ). Dieting during pregnancy was positively associated with the ‘Healthy’ and ‘Traditional’ DP and negatively with the ‘Confectionary’ DP. Body weight and shape concerns in pregnancy were associated with the ‘Healthy’ and ‘Traditional’ DP( Reference Northstone, Emmett and Rogers 30 ).
Sociodemographic
In the ALSPAC cohort, a ‘Health conscious’ DP (Table 4) was positively associated with education level and age, and was more commonly seen in women who were owner-occupiers rather than in rented accommodation( Reference Northstone, Emmett and Rogers 30 ).
As in ALSPAC, in the CANDLE cohort women adhering to the ‘Healthy’ DP were more likely older, with higher education levels and cohabiting. With regard to ethnicity, clear patterning emerged such that the ‘Processed’, ‘US Southern’ and their mixed patterns ‘Processed-Southern’ and ‘Healthy-Southern’ were more commonly consumed by African Americans, while Caucasians and women of other ethnicities tended to consume the ‘Healthy’ or ‘Healthy-Processed’ pattern( Reference Völgyi, Carroll and Hare 31 ).
In the Spanish cohort assessing DP in weeks 6, 10, 26 and 38 of pregnancy, the ‘Vegetable and meat’ pattern was positively associated with age in weeks 10 and 38( Reference Cucó, Fernández-Ballart and Sala 32 ).
Results from multiple linear regression analysis showed positive associations for age and the ‘Healthy’ and the ‘Alcohol and butter’ patterns, but inverse associations for the ‘Fast food’ pattern and the ‘Traditional meat’ pattern in the DIPP study. Education was positively associated with the ‘Healthy’, ‘Low-fat foods’ and ‘Alcohol and butter’ patterns( Reference Arkkola, Uusitalo and Kronberg-Kippila 33 ).
In the ECCAGE cohort the ‘Varied’ pattern, much like the ‘Healthy’ patterns in studies discussed above, was associated with being older and more educated. It was also associated with living with a partner, being employed and having a higher income( Reference Hoffmann, Nunes and Schmidt 6 ).
Among PHP participants, dietary guideline compliance was low; only 3·5 % of participants met all recommendations. Meeting guidelines was not associated with education( Reference Fowler, Evers and Campbell 44 ). Using the DQI-P, 56 % were classed as having sufficient DQ. In the final parsimonious model, DQ score was predicted by being a recent immigrant and being married( Reference Nash, Gilliland and Evers 35 ).
Three publications from the PIN cohort also found older age, higher education and greater income to be associated with higher DQ( Reference Bodnar and Siega-Riz 36 ). Mean DQI-P scores were higher in African-American women( Reference Bodnar and Siega-Riz 36 , Reference Laraia, Siega-Riz and Kaufman 38 ).
Another US cohort, Project Viva, assessed DQ using the Alternate Healthy Eating Index adapted for pregnancy (AHEI-P). In multivariate-adjusted models controlling for all maternal characteristics simultaneously, AHEI-P scores were positively associated with age and education. Scores initially appeared to differ by race; however, these differences disappeared upon adjustment and were found to largely stem from confounding by age and education( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 ).
Mean Mediterranean diet scores were higher in Rhea study participants who were older, more educated, married and Greek nationality( Reference Kritsotakis, Chatzi and Vassilaki 34 ). Interestingly, in another Greek study, HEI scores did not appear to be influenced by maternal age, education or income( Reference Tsigga, Filis and Hatzopoulou 42 ).
Age and education were also positively associated with DQ in a sample of low-income, un- and underinsured US women( Reference Fowles, Bryant and Kim 39 ).
A comparative study of Caucasian and Native American low-income women in recipients of federal supplemental nutrition programme assistance found no differences in DQ scores by age or income but lower mean scores in Native Americans (unadjusted for confounders)( Reference Watts, Rockett and Baer 43 ).
Individual response
In accordance with our framework, we regarded weight status before pregnancy as an individual biological and behavioural response to environmental cues. We also classed health behaviours such as smoking or physical activity as individual responses.
ALSPAC participants considering themselves ‘more active’ than their peers scored higher on the ‘Health conscious’ pattern( Reference Northstone, Emmett and Rogers 30 ).
CANDLE participants of normal pre-pregnancy weight more likely followed the ‘Healthy’ pattern, while overweight and obese more commonly followed the ‘US-Southern’, ‘Processed’ and their mixed patterns( Reference Völgyi, Carroll and Hare 31 ). In the Spanish cohort, preconception BMI was negatively associated with the ‘Vegetables and meat’ pattern in week 38 of pregnancy, while smoking was positively and physical activity negatively associated with the ‘Sweetened beverages and sugars’ pattern( Reference Cucó, Fernández-Ballart and Sala 32 ). Participants who were obese before pregnancy had 76 % greater odds of low DQ scores in the PIN cohort( Reference Laraia, Bodnar and Siega-Riz 37 ). Likewise, in Project Viva, pre-pregnancy BMI was inversely associated with DQ( Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 ). Conversely, in the Brazilian ECCAGE study no association was seen between pre-pregnancy BMI and any DP( Reference Hoffmann, Nunes and Schmidt 6 ). HEI score was negatively associated with BMI in correlational but not regression analysis in a small Greek study( Reference Tsigga, Filis and Hatzopoulou 42 ).
Smoking in pregnancy was associated with the ‘Fast foods’, ‘Traditional meat’ and ‘Coffee’ patterns in the DIPP study( Reference Arkkola, Uusitalo and Kronberg-Kippila 33 ). In the PHP study, not smoking and exercising more predicted greater DQ( Reference Nash, Gilliland and Evers 35 ) and greater mean Mediterranean diet scores in the Rhea cohort( Reference Kritsotakis, Chatzi and Vassilaki 34 ). Frequent fast-food eaters exhibited lower DQ( Reference Fowles, Timmerman and Bryant 40 ).
Environment
HEI scores were determined by place of residency; Greek women living in urban areas had increased odds of low DQ( Reference Tsigga, Filis and Hatzopoulou 42 ).
The food environment, specifically distance to outlets, emerged as a determinant. Living within 500 m of fast-food restaurants was associated with poorer DQ in univariate analysis and in the first multivariate linear regression model( Reference Nash, Gilliland and Evers 35 ). Likewise, women living 4 miles (6·4 km) or more away from supermarkets had twice the odds of low DQ( Reference Laraia, Siega-Riz and Kaufman 38 ).
Two studies investigated the social environment. After adjustment for confounders, total social capital and tolerance of diversity scores were positively associated with Mediterranean diet score. The authors offer the interpretation that social capital leads to feelings of obligation, reciprocity and self-control, which result in greater motivation to follow a healthy diet( Reference Kritsotakis, Chatzi and Vassilaki 34 ). Social support from family and friends was positively associated with DQ( Reference Nash, Gilliland and Evers 35 ).
Other factors
Anxiety was associated with the ‘Confectionary’ and depression with the ‘Vegetarian’ pattern( Reference Northstone, Emmett and Rogers 30 ) and inversely associated with DQ( Reference Nash, Gilliland and Evers 35 ). DQ was negatively associated with depression, overall and persistent stress in low-income un- and underinsured women( Reference Fowles, Timmerman and Bryant 40 ). These factors were not represented in the conceptual framework; they could build a new category or could be grouped as individual psychological responses.
Discussion
The present systematic review has synthesised seventeen publications of twelve studies on determinants of diet during pregnancy in accordance with our framework.
Factors within the category of sociodemographic determinants have been most frequently studied. Evidence consistently points to a social gradient whereby women who are older, more educated, with higher incomes or other markers of affluence more likely followed a ‘healthier’ DP or scored higher on DQ scales. A social gradient in diet has been observed in different populations and settings( Reference Darmon and Drewnowski 49 ) and in pregnant women( Reference Crozier, Robinson and Borland 50 ). However, pregnancy has been described as a period of greater motivation for behaviour change and great potential for health promotion( Reference Gardner, Croker and Barr 7 ). The fact that the social gradient in diet persists in pregnancy indicates that the health promotion potential is not used to its fullest potential, women’s motivation is not as great as expected, or that neither can overcome the wider social forces in play.
Findings regarding ethnicity are less consistent. As analyses were mostly not adjusted for confounders we find the evidence from the Project Viva cohort most convincing, where differences largely stemmed from confounding by age and education. Evidence from the reviewed studies also indicates that partnership and markers thereof such as cohabitation determine dietary intake.
Studies on individual response largely investigated health behaviours. Included studies paint a picture of a ‘behavioural’ gradient, whereby health-promoting behaviour such as adequate physical activity appears linked with higher DQ or adherence to ‘health conscious’ type patterns, whereas the opposite was seen for detrimental behaviours such as smoking. We interpret these as individual behavioural responses. The observation that diet in pregnancy ‘parallels’ other health-related behaviour before and during pregnancy corroborates with findings from different age groups and populations indicating that health-risk and health-protective behaviours ‘cluster’ together( Reference Spring, Moller and Coons 51 ). The relationship between pre-pregnancy weight and diet in pregnancy is more difficult to interpret. If body weight is interpreted as an outcome of diet this indicates that diet ‘tracks’ from preconception into pregnancy, rather than body weight being a determinant. This is supported by a prospective analysis of the Southampton Women’s Survey where DP did not change substantially upon becoming pregnant( Reference Crozier, Robinson and Godfrey 52 ).
Our review showed that pregnancy-related factors other than parity and environmental factors were less commonly investigated in studies.
The lack of studies investigating pregnancy determinants is in contrast with theoretical and empirical literature framing pregnancy as a physiologically and psychologically unique period important for health( Reference Gardner, Croker and Barr 7 , Reference Lawlor and Chaturvedi 53 ). We would have expected studies to investigate a wide range of pregnancy factors such as pregnancy intendedness, pregnancy ailments, changes in appetite and pregnancy-induced health changes for their potential influence on diet. But this was not the case; studies investigated only a few pregnancy factors other than parity. Findings on parity were inconsistent. It is possible that this is due to confounding, i.e. parity acting as a marker of age, marital status and other sociodemographic determinants, or that the influence of parity is context specific, e.g. differences in resources and support allocated to women in their first pregnancy and women who already have children.
Environmental determinants were assessed in few of the included studies. Evidence points to social support and social capital as determinants. Evidence regarding the built and food environment stems from few studies with some inconsistencies. Other facets of the environment such as medical (e.g. antenatal care) or political and economic (e.g. food policies, advertisement) were not researched. This corroborates with findings of a series of systematic reviews of determinants of diet across different age groups which also identified a lack of studies investigating macro-level environmental determinants( Reference Brug 54 ).
Psychological health emerged as a new category of determinants to add to the framework. Reviewed studies indicate that depression, stress and anxiety influence diet during pregnancy. However, we did not specifically search for these factors; these findings were thus not derived systematically. A review of psychological determinants should be conducted in order to identify where in the framework they should be placed, i.e. whether they should be regarded as a form of individual response or build an independent category.
The methodological quality of the reviewed studies raises concern. Sample size calculations were rarely reported and only nine adjusted for confounders( Reference Hoffmann, Nunes and Schmidt 6 , Reference Rifas-Shiman, Rich-Edwards and Kleinman 16 , Reference Northstone, Emmett and Rogers 30 , Reference Cucó, Fernández-Ballart and Sala 32 – Reference Nash, Gilliland and Evers 35 , Reference Laraia, Bodnar and Siega-Riz 37 , Reference Laraia, Siega-Riz and Kaufman 38 ). The assessment of dietary patterns and quality is also problematic. The reviewed studies may not be capturing exactly the same outcome (diet). Particularly DP differ between populations, places and cultural contexts and are sometimes difficult to interpret( Reference Chen, Zhao and Mao 14 ). DP are frequently derived using factor analysis, a method criticised for being based on subjective decisions( Reference Fabrigar, Wegener and MacCallum 47 ) and because results can be influenced by choice of factor loading cut-offs and rotation methods( Reference Castro, Baltar and Selem 48 ). Nutritional epidemiology has reacted to this by striving for new approaches for deriving DP such as exploratory structural equation modelling( Reference Castro, Baltar and Selem 48 ), simplified factor analyses approaches( Reference Schulze, Hoffmann and Kroke 55 ) and latent class modelling( Reference Sotres-Alvarez, Herring and Siega-Riz 56 ).
Our review benefited from an extensive literature search and quality assessment. The first step of screening and data extraction was conducted by only one reviewer. In order to counteract this potential source of bias, only articles that could be excluded without doubt (e.g. participants were not pregnant or in postpartum) were excluded based on title/abstract. Therefore 130 articles entered the second stage of screening and were read in full by two reviewers. With data extraction, any lack of clarity was resolved by discussion among reviewers. We reviewed only observational studies, which are methodologically weaker than experimental studies, because we wanted to identify the drivers of diet when women are free to choose, i.e. in real-life settings rather than experiments. Language bias is possible because all included studies were in English. Restriction to high- and upper-middle-income countries limits the generalisability of our findings.
Our framework should be seen as work in progress as this is a new research area. We recommend that more studies be conducted, particularly assessing environmental factors and pregnancy itself as a potential unique determinant. Future studies should use sound statistical techniques to overcome the issues (e.g. use of factor analysis and principal component analysis, not adjusting for confounders, lack of sample size calculations) we outlined. Once a stronger evidence base is built, it can be translated into solid public health messages and interventions.
Conclusion
Diet in pregnancy appears socially patterned and aligns along other health behaviours. Practitioners should be aware that women who are young, less educated and less affluent or who show health-risk behaviours appear to be at higher risk of poor diet in pregnancy and may require closer monitoring and advice.
Acknowledgements
Financial support: This work was carried out as part of the PhD of I.-M.D., who is currently working on the BaBi study, funded by the German Federal Ministry of Education and Research (BMBF; grant number (FKZ) 01ER1202). The primary investigators are J.S. and O.R. The BMBF had no role in the design, analysis or writing of this article. Conflict of interest: None. Authorship: I.-M.D. and J.S. formulated the research question; I.-M.D. formulated the study design. B.B. acted as second reviewer of studies for inclusion, provided valuable comments and thoroughly reviewed the manuscript. A.G. acted as a reviewer, assisted with the NOS/risk of bias ratings and commented on the manuscript. O.R. and J.S. were involved in the setting of the research agenda and thoroughly revised the manuscript. Ethics of human subject participation: Not applicable.
Appendix
Search strategies
PubMed
Filters: Humans, Adult: 19+ years: (((((((((((((determin*(Text Word) OR correlat*(Text Word) OR predict*(Text Word) OR associat*(Text Word) OR socioeconomic*(Text Word) OR socio-economic*(Text Word) OR social*(Text Word) OR econom*(Text Word) OR incom(Text Word) OR famil*(Text Word) OR household(Text Word) OR employment(Text Word) OR occupation(Text Word) OR educat*(Text Word) OR cultur*(Text Word) OR rac*(Text Word) OR ethnic*(Text Word) OR religio*(Text Word) OR marital status(Text Word) OR age(Text Word))) OR socioeconomic factors(MeSH Terms)) OR socioeconomic status(MeSH Terms)) OR marital status(MeSH Terms)) OR age factors(MeSH Terms)) OR income(MeSH Terms)) OR family characteristics(MeSH Terms)) OR cultural background(MeSH Terms)) OR household(MeSH Terms)) OR employment(MeSH Terms)) OR epidemiologic determinants(MeSH Terms)) OR statistics as topic(MeSH Terms)) AND ((((((‘diet pattern’(Text Word) OR ‘dietary pattern’ (Text Word) OR ‘food pattern’(Text Word) OR ‘meal pattern’(Text Word) OR dietary habit*(Text Word) OR food habit*(Text Word) OR meal habit*(Text Word))) OR diets(MeSH Terms)) OR maternal nutrition physiology(MeSH Terms)) OR food habits(MeSH Terms)) OR food preferences(MeSH Terms)) AND ((((pregnant(Text Word) OR pregnancy(Text Word) OR gestation*(Text Word) OR mother(Text Word) OR maternal(Text Word) OR expecting(Text Word) OR expecting(Text Word) OR parous(Text Word) OR gravid*(Text Word))) OR pregnant women(MeSH Terms)) OR pregnancy maintenance(MeSH Terms))
CINAHL and GreenFILE Library via EBSCOHOST
(determinant OR socioeconomic factor OR association OR cause OR cultural OR religion OR family OR employ) AND TX (dietary pattern OR meal pattern OR food pattern OR nutrition pattern OR diet habit OR meal habit OR food habit OR nutrition habit) AND SU (pregnancy OR pregnant OR gestation OR gravid)
MedPilot, now LVIVO
TI=(determinant OR socioeconomic factor OR association OR cause OR cultural OR religion OR family OR employ) AND TI=(dietary pattern OR meal pattern OR food pattern OR nutrition pattern OR diet habit OR meal habit OR food habit OR nutrition habit) AND TI=(pregnancy OR pregnant OR gestation OR gravid)
