The relationship between maternal food insecurity, fetal and postnatal growth: ndings from the South African National Income Dynamics Study

Background: Food insecurity during pregnancy and postnatally has been associated with adverse child health outcomes including low birthweight and stunting. There are few studies that have examined the relationships between food insecurity and child health outcomes in low and middle income countries. This study aims to ll the gap by examining the relationship between food insecurity, low birthweight and stunting in children using a longitudinal sample of South African women and their offspring. Methods: The primary exposure is a multidimensional, composite measure of household food insecurity that comprises three dimensions and six indicators. The primary outcome is a dichotomous measure of child LBW ( ≤ 2500g) The secondary outcome is childhood stunting (height for age £2SD of the median) and severe stunting (height for age £3SD of the median) in the rst ve years of life. We used both unadjusted and adjusted regression models. Source data for the analysis came from the population-based, longitudinal data collected for wave one and three of the National Income Dynamics Study (2008 and 2012). Results: Birthweight data were available for 1381 children and stunting data were available for 1444 children. The prevalence of LBW in the sample was 13.64 %. The prevalence of stunting and severe stunting was 16.46 % and 11.87 %. The composite measure was not associated with low birthweight but was associated with underweight maternal BMI, a risk factor for low birthweight. Household dietary diversity, food expenditure below the Stats SA poverty line and the composite measure was associated with stunting and severe stunting in bivariate analyses. The measure was associated with severe stunting in the nal model. Conclusions: The nding that household food insecurity is associated with severe stunting for children in the rst ve years of life highlights the importance of adequate food and nutrition for vulnerable children in South Africa. Future studies can elucidate the relationship between maternal food insecurity, low birthweight and other child outcomes. Policy interventions that focus on early childhood growth through alleviating food insecurity and addressing other drivers of malnutrition such as improving maternal education and dietary diversity can break the cycle of poor growth outcomes.


Introduction
Pregnancy and the postpartum period are vulnerable times for women in both high and low and middle income (LMIC) countries. In LMIC many pregnant women live in poverty and may experience a number of stressors including intimate partner violence, economic hardship and food insecurity (1)(2)(3)(4)(5)(6).
Observational epidemiological evidence from both high-and low-middle income countries on food insecurity and child health outcomes has yielded con icting ndings. A population-based case control study from the United States of America (USA) by Carmichael and colleagues (7) suggested an increased risk of birth defects among women who were food insecure during their pregnancies . The authors of the study hypothesized that food insecurity is associated with maternal distress as well as compromised nutrition, both of which are associated with the etiology of birth defects (7). Cross-sectional data from Bangladesh examined the relationship between food insecurity and infant birth size as a proxy for birthweight (8). While this study did nd a signi cant relationship between food insecurity and infant birth size it has a number of limitations and thus the results need to be interpreted with caution These included the subjective nature of the outcome (birth size) as well as a ve-year recall period. Among a sample of older children and toddlers, a USA based retrospective cohort study linked household food insecurity to an increased risk of poor health and hospitalization in early childhood (9).
These ndings of previous studies described above suggest that food insecurity during antenatal or early childhood periods may be related to child health outcomes. However, several key questions remain for the eld.
Women living in LMIC or in high income countries with conditions of inequality are likely to experience food insecurity during their pregnancies and while some may be malnourished, this is not true for all (10)(11)(12). Food insecurity spans a spectrum from anxiety about household food supply to actual hunger in the household and little is known of how pregnant women in LMIC may experience this spectrum of food insecurity. The authors of this study sought to explore the relationship between household food insecurity during the antenatal period with birthweight and stunting in the rst ve years of life using data from a longitudinal population-based panel survey in South Africa.

Methods
The South African National Income Dynamics Study (SA-NIDS) is a nationally representative panel survey of households in South Africa (18). The survey which began in 2008 is conducted by the Southern Africa Labour and Development Research Unit based at the University of Cape Town, Cape Town, South Africa. To date, data from the rst ve waves of NIDS are available with Wave 5 concluding in 2016. The study utilizes a strati ed twostage cluster sample design to randomly select 400 of Statistics South Africa's 3000 primary sampling units (PSUs) for inclusion in the surveys (18).

Ethics Approval
The SA-NIDS was approved by the Ethics Committee of the Commerce Faculty, University of Cape Town, and the de-identi ed datasets are publicly available. Ethics approval for this secondary analysis was obtained from the University of the Witwatersrand Research Ethics Committee (protocol number M1909101).
The components of the secondary analysis on which this paper is based are described below: The current paper utilized data from wave 1 and 3 of NIDS. The authors obtained household level food security data and individual maternal characteristics from wave 1 household and adult questionnaires collected in 2008.
During data collection, the majority of women (77.7%) were in the antenatal period while 11% reported being pregnant and 11.3% had recently given birth with infants having a mean age of 3.9 months. The data from women who became pregnant between wave 1 and 2 of the study were matched with the data of their children from wave 3. Information on children was obtained from the responses to the child questionnaires which were answered on behalf of children by their mother.

Inclusion and exclusion criteria
We limited our analysis to women aged between 15 and 44 because younger adolescents and older mothers are both at increased risk of obstetric complications and LBW and may differ systematically from the population of interest (19). Data on children were limited to children born between January 2008 and December 2011, the commencement of data collection in wave 1 and the conclusion of data collection in wave 2. Children were aged between one and ve years at the time of data collection with a mean age of 38 months.

Food Insecurity
The primary exposure during the antenatal period was a composite household food insecurity score between 0 and 6, that used a methodology of multidimensional poverty measurement developed by Alkire and colleagues, in calculating an index of food insecurity (13). This adapted measure of food insecurity includes subjective and objective indicators associated with the experience and outcomes of food insecurity (14). A previous article by Ryan and colleagues also included child anthropometry in the measure but given that this was our outcome of interest we removed this indicator to avoid collinearity (14). Each indicator is assigned a value of 0 or 1, with the nal score being a minimum of 0 (food secure) and a maximum of 6 (severely food insecure). Given the low percentage of underweight women, households were classi ed as food insecure if they scored more than one in the food quality and food anxiety domains (i.e. 2 out of 3 total domains). The food insecurity score was examined as a continuous and binary (domain insecure) variable. This measure covers three domains (anxiety at the household level, food quality at the household level and physical consequences at the individual level) and includes six indicators (the subjective indicators of adult and child hunger and household food su ciency and the objective indicators of dietary diversity, proportion of household expenditure on food and maternal underweight BMI) (14). Adult and child hunger was de ned as households where an adult or child never or seldom went hungry versus households where an adult or child sometimes, often or always went hungry in the past year. Household food adequacy was de ned as households who reported that food consumption was more than adequate or just adequate for household needs versus less than adequate.
A household dietary diversity score (HDDS) based on 30-day recall was calculated using the Food and Agriculture Organization (FAO) guidelines (15). The HDDS is comprised of 32 individual food types and 12 different food groups. There is no standardized cutoff to describe low dietary diversity, but one recommendation is to use the mean HDDS score as a cutoff (15). The mean dietary diversity of the sample was 9.06 which indicates a fairly high number of households with diverse diets. Household scores are generated on a whole number basis and the authors therefore used a cutoff of 9 for low dietary diversity. For household food expenditure, a cutoff of expenditure above 0.6 was used to de ne a household as food insecure, as recommended by the FAO. Although measurement that utilizes household expenditure data is less exact than food consumption surveys, it is fairly accurate, and has similar estimates of food insecurity (16). A cutoff of BMI<18.5 was used to classify women as underweight per the WHO growth standards (17).

Covariates
In addition to the primary exposure of the household food security score, we included variables associated with LBW and stunting in the literature in our models. These included both distal and proximate characteristics such as maternal age, height, BMI, smoking status, depression status measured using the CES-D scale (18), employment, years of education and self-reported health (19)(20)(21)(22)(23)(24). At the household level we included household size, individual food security domains, per capita food expenditure below the Stats SA Poverty level of R274 and geotype (urban, rural or traditional areas) as well as province of residence. Child characteristics included gender, child's age in months and social grant status (25,26).

Primary Outcomes
The primary outcome is a dichotomous measure of child LBW (£2500g). The secondary outcome is childhood stunting (height for age ≤2SD) and severe stunting (height for age ≤3SD) in the rst ve years of life for children born to women between wave 1 and 2 of the NIDS study. Children's Z scores were calculated using the WHO child growth standards (17). Maternal anthropometry was collected in wave 1 of the study, while child birthweight and anthropometry were recorded in wave 3. Birthweight information was obtained from the Road to Health Card (RTHC), a child health information booklet provided by the Department of Health when women give birth (27).

Statistical Analyses
There are several steps to the analysis: First, initial unadjusted bivariate analyses to explore covariates associated with LBW, stunting and severe stunting were conducted. Second, birthweight and height for age scores were analysed both continuously and as binary variables. Third, forward stepwise logistic regression was employed to identify potential explanatory variables associated with LBW, stunting and severe stunting. We conducted logistic regression for the nal LBW model and multivariable logistic regression for the nal stunting model. The latter model incorporated stunting and severe stunting. Variables included were identi ed from the bivariate analyses and exposures with a strength of association of p<=0.05 were then selected. These variables were added to the model one at a time, with those showing the strongest association added rst. Those that were not signi cant in the stepwise model (p>0.05) were removed. Last, post-strati cation weights were adjusted to the age-sex-race distribution produced by Statistics South Africa based on population estimates. Analyses were based on a dataset with imputation values generated by the NIDS data team (18). All analyses were conducted using Stata 15 (Stata Corporation, College Station, TX).
A series of logistic regression models were run to compute odds ratios (ORs) with 95% con dence intervals (CI) with LBW as the primary outcome. This model aimed to determine the strength of the associations between food insecurity and other sociodemographic risk factors and LBW. Separate regression models were run to explore associations with stunting and severe stunting in children aged 0 -60 months. Signi cant variables were combined in a nal multiple logistic regression model that included both stunting and severe stunting and nonsigni cant variables excluded.

Sample characteristics
We matched the data from 1685 mothers with the data from their children who were born between Wave 1 and 2 (2008)(2009)(2010)(2011). A large number of children were missing anthropometry and birthweight data in wave 2 of data collection and authors therefore matched this data from wave 3 (2012). Of this sample, 1381 (82%) of children had information on birthweight and 1444 (86%) of children had anthropometry information. Both birthweight and anthropometry were available for 1190 (71%) children. The mean maternal age was 25.2 years and the majority of women (86%) belonged to the black African population group and were unemployed or not economically active (72%). Most women were unmarried (68%) while 31% were married or lived with a partner. Over half (64%) of women had given birth before, while the remainder became parents for the rst time between wave 1 and 2.
Additional information on maternal characteristics is presented in Table 1.
Information on household food adequacy and food expenditure was available for the full sample with birthweight data but between 5% and 12% of data on household dietary diversity, adult and child hunger and maternal BMI were not available for all mother-child pairs. Of the children with birthweight data, 164 (12%) of mothers were missing anthropometry data. This limitation is further discussed at the end of this article. There was a total of 1299 complete food security scores for the sample. Scores varied between 0 and 6 with a mean of 1.40. With a cutoff of 2, 79% of households were classi ed as food secure while the remaining 21% were food insecure. Only one household (0.08%) had maximum scores for all indicators with a total score of 6, but 15% of households had scores of 4 or above.
Another way to classify food insecurity using the score is to look at the scores across domains (14). Given the low number of underweight women in this sample, we classi ed households as domain insecure if they scored above 1 in the food quality and food anxiety domains (i.e. 2 out of 3 total domains). A total of 30% of households were food insecure at the domain level. This may be a better indicator of food insecurity as it produces more variation in the exposures than the sum of the individual indicators and there are more observations for the full sample.   Anthropometry information was available for 1444 children. The prevalence of stunting and severe stunting was 16.46 % and 11.87 % respectively with 238 children classi ed as stunted and 171 as severely stunted. BMI information was missing for 64 children but a total of 21% of children were overweight or obese (BMI>2 SD) while 6% of children were wasted (BMI<-2 SD). Of the children who were overweight and obese, a total of 120 (41%) were simultaneously stunted or severely stunted. Anthropometric comparisons are somewhat limited by the small number of children in the 0-24 month category. Levels of severe stunting and overweight and obesity were highest among children in the 25-48 month category and subsequently decreased as children aged.

Bivariate analyses
In unadjusted bivariate analyses, households with per capita food expenditure below the Stats SA poverty line (R274) and low dietary diversity were the two food security indicators most strongly associated with both stunting and severe stunting. Low dietary diversity increased the risk of stunting by 65% and more than doubled the risk of severe stunting. Children from households food insecure at the domain level were 42% and 81% more likely to be stunted and severely stunted. When analysed as a continuous variable, the food security score was associated with 18% and 29% increases in stunting and severe stunting respectively. Male children were more likely to be severely stunted than female children although there were no gender differences among stunted children. Each additional kilogram of birthweight was associated with a 33% reduction in stunting but birthweight was not associated with severe stunting. Increased years of maternal education, greater maternal height and child's age in months reduced the risk of both stunting and severe stunting in this sample.
Variables associated with stunting and severe stunting were combined in a nal multiple logistic regression model. To avoid collinearity between food security scores, domains and individual indicators the authors only included the composite food security score in the nal model with maternal and child characteristics from the bivariate analyses including maternal height, years of education and child's age in months, gender and birthweight. This model includes only variables that remained signi cant for either stunting or severe stunting. Lower Maternal height was the only variable that remained signi cant for both stunting and severe stunting in the nal model while years of maternal education and the food security score were signi cantly associated with severe stunting with maternal education offering a protective effect. Birthweight data were obtained for a total of 1381 children. Mean birthweight was 3109 grams with a SD of 554 grams. The prevalence of LBW in the sample was 13.64 % (a total of 188 babies). We included babies whose birthweight was recorded at exactly 2500 grams in our de nition. In initial bivariate analyses the maternal characteristics associated with LBW included elevated maternal blood pressure (p= 0.035) and underweight maternal BMI (p=0.04). Female infants were also 38% more likely to be LBW but this association lost signi cance when we adjusted for the panel design of the study. When analysed as a continuous variable, each unit increase in the food security score increased the risk of maternal underweight BMI by 66% (p=0.000) although this nding is not included in the table below. Variables that remained signi cant and were included in the nal model were underweight maternal BMI and elevated maternal blood pressure. Stage 1 maternal blood pressure was bordering signi cance with a p value of 0.053. Underweight maternal BMI was the only food security indicator associated with LBW in this sample.

Discussion
This study has revealed important ndings about the relationship between food insecurity and child outcomes in the rst ve years of life. In summary we found that the most prevalent indicators of food insecurity were low dietary diversity and self-reported household food inadequacy. The prevalence of LBW was 13.64 % while stunting and severe stunting were 16.46 % and 11.87 % respectively. The composite food security score was not directly associated with LBW, but women who were domain food insecure were 2.44 times as likely to be underweight than their counterparts, a risk factor for LBW. Elevated maternal blood pressure was also associated with LBW. Among severely stunted children, food expenditure below the food poverty line, low household dietary diversity and households that were domain insecure were the food security indicators most strongly associated with severe stunting. Among stunted children, food expenditure below the food poverty line was most strongly associated with stunting. For both stunted and severely stunted children, increased maternal height and education decreased risk as well as each additional month of child's age. In the nal multiple logistic regression model, each unit increase of the food security score increased the risk of severe stunting by 24% while maternal education and height reduced the risk.
In 2015, a quarter of the world's LBW babies were born in sub-Saharan Africa and risk factors included extremes of maternal age, preterm birth, infectious diseases and maternal nutritional status (24). Birth attendants' recording of a newborn's weight as 2500g when s/he weighs slightly below to avoid classifying as LBW is common practice and causes underestimation of the prevalence of the problem (24). The authors therefore used a cutoff of birthweight £2500g. The prevalence of LBW in the sample of 13.64% is closely aligned with recent cohort studies of birthweight in South Africa (22,28).
Although the food security indicators were not directly associated with LBW or stunting in this sample, food insecurity may still in uence these outcomes indirectly by increasing maternal risk to give birth to a LBW baby who may be subsequently more likely to experience compromised growth. Underweight maternal BMI was a risk factor for LBW although this nding is not included in Table 5. Only 5% of women in the sample were underweight while 50% were overweight or obese, indicating that food insecurity among the majority of South African women during the antenatal period is not severe enough to cause malnutrition or wasting. Among these women, it appears that the most food insecure were more likely to be underweight or normal weight. The association between LBW and underweight maternal BMI is supported by a systematic review and meta -analysis of over one million women that noted underweight women were 48% more likely to give birth to a LBW infant (29). Another factor associated with LBW included elevated maternal blood pressure. Women with elevated blood pressure (systolic between 120 and 129 mm Hg) were more likely to give birth to a LBW baby than women with stage one or stage two hypertension which is somewhat surprising although stage 1 hypertension was bordering on signi cance in the nal model with a P value of 0.053. This may be due to other risk factors for LBW such as preterm birth that were not included in the data. Hypertension is a risk factor that has been linked to food security, obesity and increased BMI in developing countries and may be partly attributable to a poor quality diet high in saturated fats and sugar (30)(31)(32)In this sample of women, increased household dietary diversity decreased BMI as well as the risk of overweight and obesity although these results are not included in the tables. These ndings highlight the complex links between food insecurity, poor quality diet and maternal overweight and obesity occurring concurrently with stunted children in the same household (32)(33)(34).
The total percentage of children in this sample who were stunted or severely stunted was 28.3 %. These gures are comparable to those estimated by two large population-based surveys, that found a stunting prevalence around 27% in children under the age of ve in South Africa (35,36)The nding that multiple food security indicators as well as the food security score were associated with both stunting and severe stunting highlights the deprivation that many South African children experience on a daily basis. While food insecurity status is not static and subject to change, these ndings indicate that chronic household food insecurity during the antenatal period likely extends into the early years of childhood.
Studies from Kenya, India and Bangladesh have also noted the association between low dietary diversity and stunting in early childhood (37)(38)(39). Mean dietary diversity in this sample was 9.24 which indicates a fairly high number of households with diverse diets. However, the measure had a thirty day recall and this may have introduced some bias into the nal scores. This is further discussed in the limitations section. In bivariate analyses of this sample the prevalence of stunting decreased as children grew older, with each additional month of age decreasing the risk of stunting by 4%. This is not surprising as previous studies suggest that stunting tends to peak around 24 months and subsequently decrease (40)(41)(42). However, among severely stunted children in this sample, age did not have an effect. Our nding that shorter mothers are more likely to have stunted children corroborates evidence from previous studies that inherited maternal characteristics (often arising from de cient fetal and early childhood environments) perpetuate stunting among their children (43)(44)(45).
The nding that increased years of maternal education has a protective effect against severe stunting has been observed in multiple studies from Sub-Saharan Africa as well as Latin America (25,(46)(47)(48)(49)(50). An impact study of the child support grant (CSG) did not observe an increase in children's height for age scores for all children receiving the child grant, but for children whose mothers had more than eight grades of schooling, the grant increased HAZ by 0.19 standard deviations (51). This may be attributable not only to increased income for these mothers but also to better knowledge and practices around childcare and infant feeding practices as education increases. Investments in education for girls and women can empower them to make different choices for their children and improve the health of subsequent generations.
Birthweight and LBW is often a predictor of future stunting (22,24,52) and we did nd an association with LBW and stunting in unadjusted bivariate analyses but this did not remain when we adjusted for the panel design of the study. This is somewhat surprising as it suggests that the most severe linear growth retardation in this sample is primarily a product of a de cient growth environment and not related to prenatal pathology or inherited birth characteristics such as weight at birth.
A recent review of the evidence base around stunting noted that linear growth retardation is associated with (but does not cause) delayed cognitive development and educational outcomes, lower earnings in adulthood, and an increased risk of chronic diseases (43). This nding is corroborated by a recent analysis of NIDS data by Casale and colleagues that noted lower educational attainment, and increased likelihood to fail a grade among stunted children (41). It is important to note the inherent problems with the de ned cutoff of -2SD as nothing changes just above or below this number and it does not have a clinical or biological basis. Furthermore, the number of stunted children will be vastly underestimated if the stunting prevalence is greater than 2.5% as the distribution of height for age scores is completely shifted (43). In spite of these limitations, stunting remains an important child health and growth indicator as it has important long-term growth and development implications. In this sample, 163 (11.3%) of children who were stunted had overweight or obese mothers. This phenomenon, whereby stunted children and overweight or obese mothers reside within the same household highlights the double burden of malnutrition that are particularly prevalent in countries undergoing a nutrition transition (34).
In South Africa, the prevalence of stunting has remained virtually the same over the past two decades, affecting a quarter of children under the age of ve years (53). In spite of national increases in many food security indicators and a high proportion of caregivers accessing the the child care grant, the prevalence of stunted growth continues to uctuate around the 25% level (25). An analysis that explored this apparent paradox noted that social grants are insu cient to end the intergenerational cycle of poor health and malnutrition, and the amount of the child support grant does not afford a basic food basket (25).

Limitations
Our study has several limitations; data on gestational age were not available and preterm birth (a leading cause of LBW) could therefore not be included in our analyses. Our study did not use a validated measure for food insecurity but constructed a composite measure from several different indicators based on published literature (15)(16)(17)Our study also assumed that household food security status remained more or less constant in the period between wave 1 and 3. The HDDS in this sample was relatively high with a mean score of 9.06 but it is important to bear in mind that the questionnaire had a thirty-day recall period and this may have falsely in ated the results as the DDS does not measure the frequency or amount of consumption. For example, a household that consumed red meat once in the past thirty days would still receive a score for that food group. Households with higher per capita food expenditure were more likely to be missing maternal anthropometry information. These results are in keeping with the analysis of the NIDS wave 1 health data that noted missing anthropometry data for 11% of women and observed that respondents living in urban areas and wealthier households are more likely to have missing anthropometric measurements (33). This may have biased our sample to include fewer women who are food secure and could have led to an underestimation of the magnitude of difference in outcomes between children in food insecure and food secure households. In addition, information on maternal HIV status was not included in the study and this may be an important confounding factor for birth outcomes as well as food security (54,55).

Conclusion
This study utilized data from a longitudinal national survey to examine the relationship between household food insecurity and child health outcomes. We found that household food insecurity, low dietary diversity and food expenditure below the food poverty line were associated with stunting and severe stunting in the rst ve years of life as well as maternal underweight BMI (a risk factor for LBW).
Our ndings indicate that stunting and severe stunting in the rst ve years of life are more likely a consequence of food insecurity and other aspects of a de cient growth environment than an outcome of low birthweight or other maternal exposures during the antenatal period. Our study also nds that dietary diversity is a predictor of severe stunting and an effective proxy for food insecurity in studies that do not utilize a validated food security measure. However, improved food security alone is not a panacea for the poor growth outcomes that negatively impact so many South African children. The proximal drivers of stunting and malnutrition are complex and encompass the broader issues of social inequality in South Africa that include high levels of poverty, food insecurity, and poor access to healthcare and sanitation services (53). At the time of writing, the world and South Africa were in the early stages of the Covid-19 pandemic and the indirect health and economic consequences were already being felt among South African children with decreased rates of hospital admissions, a decrease in routine vaccinations and an expected increase in childhood malnutrition (56,57) . Given the multiple vulnerabilities of the mothers in this study (i.e. 23% adolescents and 73% unemployed), the authors assume that food insecurity and its adverse consequences for early childhood health will be magni ed by the current crisis.
It will be a useful avenue for future research to try and elucidate the magnitude of effect of the various drivers associated with stunting. Interventions that address these other drivers of malnutrition, such as maternal education, childcare, breastfeeding and hygiene practices, as well as dietary quality are needed to ensure that South Africa children can reach their full growth potential. participants, and only proceeded with interviews once this process was complete. Adult caregivers consented on behalf of minor participants and completed questionnaires on their behalf.

Consent for Publication
Not Applicable Availability of data and materials Deidenti ed data from the South African National Income Dynamics Study (NIDS) can be accessed via the NIDS website (http://www.nids.uct.ac.za/). Alternatively, data that contains con dential information can be accessed via an application process with Data rst at the University of Cape (www.data rst.uct.ac.za).

Competing interests
The authors declare that they have no competing interests.