Effects of vitamin and mineral supplementation during pregnancy on maternal, birth, child health and development outcomes in low‐ and middle‐income countries: A systematic review

We included the following study designs: Randomized controlled studies (RCTs), where participants were randomly assigned, individually or in clusters, to intervention and comparison groups. Cross‐over designs were eligible for inclusion. Quasi‐experimental designs, which include: – Natural experiments: studies where non‐random assignment is determined by factors that are out of the control of the investigator. One common type includes allocation based on exogenous geographical variation. – Controlled before‐after studies (CBA), in which measures were taken of an experimental group and a comparable control group both before and after the intervention. We also required that appropriate methods were used to control for confounding, such as statistical matching (e.g., propensity score matching, or covariate matching) or regression adjustment (e.g., difference‐in‐differences, instrumental variables). – Regression discontinuity designs; here, allocation to intervention/control is based upon a cut‐off score. – Interrupted time series (ITS) studies, in which outcomes were measured in the intervention group at least three time points before the intervention and after the intervention.

Reviews were excluded.

Participants were healthy (i.e. non‐diseased) pregnant women of any age and parity living in LMIC. LMIC will be defined by the World Bank Group at the time of the search for this review. Though our aim was to include healthy pregnant women, the prevalence of micronutrient deficiencies is high in these settings, indicating that women are likely to have one or more micronutrient deficiencies at baseline; women were not excluded on this basis. Studies that include only a subset of eligible participants were retained as eligible, but were only included in analysis where data had been disaggregated appropriately for use.

The following interventions targeting pregnant women were included, and were analysed separately:Single micronutrient supplementation (calcium, vitamin D, iodine, folic acid, iron, vitamin A, zinc, vitamin B12) compared to placebo (Supplementation may take the form of tablets, drops, syrup or powder)Iron folic acid supplementation compared to folic acid alone or placeboVitamin D and calcium supplementation compared to placeboMMN compared to iron folic acid supplementation or placebo: For MMN, studies that use fewer than 3 micronutrients in its composition were excluded (Haider & Bhutta, 2017; Kawai et al., 2011)LNS compared to MMN or placebo

For logistical reasons, we did not include every vitamin and mineral. Interventions were chosen based on relevance (i.e., most prevalent nutritional deficiencies) and data availability when considering the LMIC context.

There were no restrictions regarding: i) the duration of exposure to the intervention, ii) the provider of the intervention, iii) the frequency of the intervention (e.g. daily or intermittent supplementation), or iv) the food vehicle utilized for LNS interventions. We included studies where co‐interventions (e.g. education) were provided for both the intervention and the comparison groups.

To be included within this review, studies measured at least one of the following primary and/or secondary outcomes. We looked at maternal, fetal, neonatal and child health and nutrition outcomes that will help to inform related policy and practice. For simplifying, we grouped all secondary outcomes of interest by these domains. Unless otherwise specified, all outcomes listed were dichotomized (yes/no). We used mean and standard deviation (SD) to report all continuous outcomes (maternal biochemical status, newborn anthropometry, newborn/child biochemical status). Outcome definitions can be found in brackets below. International Units (IU) were used for all maternal outcomes whereas z‐scores were used for child outcomes, because z‐scores are adjusted for age.

Maternal Mortality (death while pregnant or within 42 days of pregnancy termination)Anemia/iron deficiency anaemia in third trimester of pregnancy [WHO, 2011]–Non‐anemic: > or equal to 110 g/L–Anemic: <110 g/Llow birthweight (<2500 g)perinatal mortality (stillbirths and deaths <or equal to 7 days)

There was no minimum duration of follow up.

Other than the LMIC inclusion criteria, there were no restrictions regarding study setting.

Any post hoc changes to eligibility criteria or outcomes studied were aligned with the review objectives and were clearly stated with reasons justified.

The search was run in the following databases, selected based on their applicability to the subject material: CAB Abstracts CINAHL Cochrane Central Register of Controlled Trials (CENTRAL) Embase International Initiative for Impact Evaluations (3ie) LILACS (Latin American and Caribbean health sciences literature) Medline POPLINE Web of Science WHOLIS (WHO library database)

The vast majority of included studies were randomised or cluster‐randomised controlled studies that followed our inclusion/exclusion criteria, as listed above. For example, in a study published by Christian and colleagues (Christian et al., 2003), pregnant Nepalese women were cluster‐randomised to 1 of 5 groups: i) daily supplements of vitamin A (control), ii) daily supplements of vitamin A + folic acid, iii) daily supplements of vitamin A + folic acid + iron, iv) daily supplements of vitamin A + folic acid + iron + zinc, or v) daily multiple micronutrient supplements (including vitamin A) from early pregnancy to 72 hours postpartum. Outcomes assessed included infant anthropometrics (birth weight, length, and head circumference), and low birthweight (<2500 grams).

In order to take into account potential sources of dependency, we grouped studies in terms of their location, population, the programme that is being evaluated (if applicable), and intervention type to ensure that there is no double counting of evidence when synthesizing results across studies. Where there were multiple papers that described the same study, these papers were combined and coded as a single study.

For studies that include multiple intervention arms, we selected one pair (intervention and control) that satisfied the inclusion criteria of the review and excluded the rest. If >2 intervention groups met the eligibility criteria, then these groups were combined into a single pair‐wise comparison group and data was disaggregated into corresponding subgroups, or these arms were separated into different forest plots to ensure that there was no double counting of participants. Where multiple effect sizes from the same study and intervention‐control comparison were eligible, we chose to include the estimate with the larger sample size, though this occurred rarely.

Two independent reviewers performed title and abstract screening using specified inclusion/exclusion criteria. Where not enough information could be gleaned from the title alone, then abstracts were screened in order to determine eligibility for full text screening. All full texts were then screened in duplicate, with application of the same inclusion/exclusion criteria. A third reviewer resolved any disagreements. Both title/abstract and full text screening were done using Covidence, a web‐based software platform for systematic reviews. We assessed inter‐reviewer reliability/agreement by checking the number of conflicts in the Resolve Conflicts page following each stage of screening.

For all included studies, we extracted data into a standardized data abstraction form that was comprised of a general study information sheet and a quantitative outcomes sheet. The data abstraction form was piloted before it was finalized. While all arms of a study were described in the tables of included studies, data was extracted and reported on only for those arms that met review criteria. All data abstraction was performed in duplicate. Coders were trained in systematic review methods, and data abstraction was cross‐checked with primary study data for accuracy by the team lead.

Each general study information sheet contained the following:General study information: authors, publication year, language of study, study designStudy setting: World Bank region, country, World Bank income level, city/town, urban/urban slum/rural/mixed setting, duration of data collection, date of data collectionStudy population: sample size recruited, sample size analysed, male/female/mixed (%), age range of participants, mean/median age of participants, description of participants (i.e. inclusion/exclusion criteria applied to recruitment)Intervention characteristics: type of intervention, food vehicle utilized (where applicable), duration of intervention, level of delivery, unit of randomisation (where applicable), dose of micronutrient(s) provided, frequency of provision (i.e. daily, weekly, etc.), duration of follow up, attrition rateProgrammatic indicators (based on the WHO/CDC logic model (De‐Regil et al., 2014)): policies, production, delivery strategies, quality control, behaviour change communication strategies, access and coverage, knowledge and appropriate useFunding source of programme (where applicable)Quality assessment (see section below: critical appraisal of studies

Each quantitative outcome sheet contained the following: Subgroup (if applicable) Subgroup sample size Outcome type (based on outcomes listed above) Outcome units Outcomes: – Outcome measure treatment group – Outcome measure comparison group – Standard deviation Effect size: – Effect measure (specify type); unadjusted and adjusted – 95% confidence interval (CI) – P‐value of effect measure – Standard error (SE) or standard deviation (SD) or t‐statistic

We critically appraised individual studies using the Cochrane Effective Practice and Organisation of Care (EPOC) guidelines for randomised studies, non‐randomised studies, controlled before‐after studies, and interrupted time series (ITS) studies. EPOC guidelines include the following standardized criteria for assessing bias of randomised, non‐randomised, and controlled before‐after studies [Cochrane Effective Practice and Organisation of Care (EPOC), 2017]: Random sequence generation Allocation concealment Baseline outcome measurements similar Baseline characteristics similar Incomplete outcome data Knowledge of the allocated interventions adequately prevented during study Protection against contamination Selective outcome reporting Other risks of bias (e.g. bias in measurement: validity and reliability of the measures used)

For ITS studies, the following criteria was considered [Cochrane Effective Practice and Organisation of Care (EPOC), 2017]: Intervention independent of other changes Shape of intervention effect pre‐specified Intervention unlikely to affect data collection Knowledge of the allocated interventions adequately prevented during study Incomplete outcome data Selective outcome reporting Other risks of bias (e.g. bias in measurement: validity and reliability of the measures used)

For EPOC rating schemes for randomised studies, non‐randomised studies, and controlled before‐after studies please see Table 2 and for interrupted time series studies, see Table 3.

In addition, the Cochrane risk of bias (ROB) tool (Higgins & Green, 2011) was used for randomised studies, including cluster‐randomised studies and step‐wedge designs. The ROB tool used the following criteria for assessment of bias. Of note, we assessed performance and detection bias separately.Selection bias: random sequence generation and allocation concealmentPerformance bias: blinding of participants and personnelDetection bias: blinding of outcome assessmentAttrition bias: incomplete outcome dataReporting bias: selective reportingOther sources of bias

All risk of bias assessments were performed in duplicate and supportive evidence for all ROB judgements was documented. A third reviewer resolved any disagreements. An overall score was not provided.

We converted data for each outcome into the same format (e.g. means and standard deviations for continuous data), including appropriate conversion of scales such that an increase/decrease always indicated improvement or deterioration of an indicator. In the case that included studies had data that were reported in a not 'usable' way (i.e., data cannot be pooled with other data), we retained the study as eligible but restricted it from further analysis.

We analyzed dichotomous and continuous outcomes separately. For dichotomous outcomes, results were presented as summary risk ratios (RRs) with 95% CIs, whenever possible, in order to compare risk of the outcome between intervention and control groups. When including incidence data, we combined risk ratios (events per child) and rate ratios (events per child year) because of their similar interpretation and scale. We presented continuous outcome data as either a mean difference (MD), if outcomes were measured on the same scale, or a standardized mean difference (SMD), if outcomes were measured on different scales, with 95% CIs. Both change from baseline scores and final measurements (for RCTs only) were eligible, and were pooled where there is meta‐analysis with MD (i.e., scales are the same and measurements are in the same unit) Higgins & Green, 2011. We carefully considered reporting of the appropriate means and standard deviations (either of final measurements or of changes from baseline) if both change and final values were used in one meta‐analysis. We did not combine final values and change scores as SMDs because the standard deviation in this case reflected differences in measurement reliability. Where it was necessary to combine measures of treatment effects with SMDs (for example, in the case of child development outcomes), we used change scores as opposed to endline values, but only where authors had appropriately reported the SD of the change score.

All interventions and, within those interventions, outcomes were meta‐analysed separately. We also meta‐analysed RCTs and quasi‐experimental studies separately.

Special attention was given to cluster‐randomised studies; this was to ensure that clustering was appropriately accounted for within the analysis of the primary study, such that study precision was not over or under‐estimated within our analysis. If necessary, we adjusted effect estimates of cluster‐randomised studies using the mean cluster size (M) and the intra‐cluster correlation coefficient (ICC), which quantified the extent to which data from the same cluster were correlated [design effect = 1 + (M‐1) ICC]. The design effect was then used to adjust the study data such that a study was reduced to its effective sample size. We did not make any adjustments if authors appropriately adjusted for cluster design already.

Randomized and non‐randomised studies with contemporaneous comparison groups were analysed separately, but were pooled if differences in findings were not statistically significant. We analyzed and reported findings from controlled before‐after and ITS study designs separately.

Where data was incomplete or in a form that could not be converted with the information available, we contacted the corresponding author for clarification or to obtain missing data. If authors have accounted for missing data (i.e. multiple imputations), we used the adjusted data within our analysis.

Statistical heterogeneity was assessed using Tau², I² and significance of the Chi‐square test; we also assessed heterogeneity visually using forest plots. Based on prior theory and clinical knowledge, we expected clinical and methodological heterogeneity in effect sizes in this literature. Therefore, we attempted to explain any observed statistical heterogeneity using subgroup analysis (see below).

If the number of studies was sufficient (>10), funnel plots were used to visually assess publication bias. This kind of bias is unlikely if data forms a symmetric inverted funnel shape around the mean effect estimate. In addition, we performed Egger's test to determine funnel plot asymmetry.

Statistical analysis was carried out using Review Manager 5.3 and Stata. We followed intention to treat (ITT) analysis for RCTs. We reconstructed the data to create an ITT analysis where authors reported a per protocol analysis.

Random effects meta‐analysis was used to account for any expected heterogeneity in interventions, comparisons, outcomes, or settings within the studies included in a given synthesis. Where meta‐analysis was deemed inappropriate due to substantial methodological or statistical heterogeneity between studies, we summarized the findings of the included studies in narrative or table form.

The generic inverse‐variance approach was used for both dichotomous and continuous outcomes, such that the study weights were adjusted according to the variance of the effect estimate (i.e. the larger studies with smaller standard error were given more weight than smaller studies with larger standard error). For random effects analyses, the DerSimonian and Laird method was applied to incorporate a measure of variation (Tau²) among intervention effects from different studies.

We used raw summary estimates to construct meta‐analyses from RCTs and adjusted estimates to construct meta‐analyses from observational studies. We interpreted overall effect estimates that had an associated p‐value <0.05 as statistically significant, but also commented on those effects where the upper or lower confidence interval has just crossed the line of no effect. In the case of the latter, where the confidence intervals fell between the line of no effect and 0.06, the effect estimates were interpreted with caution as possibly significant (i.e. may have or may have not had an effect). Where confidence intervals were >0.06, effect estimates were interpreted as not significant.

We also reported non‐significant findings. Where possible, interaction tests were used to determine if there was a relevant difference in effect across sub‐groups. We based the conclusion that an intervention was effective in one subgroup but not another on a direct test of the mean difference between two groups (i.e., with meta‐regression).

We used the GRADE tool to assess the body of evidence for selected outcomes for which a meta‐analysis was conducted. We chose the following outcomes: maternal mortality, maternal anaemia, low birthweight, and perinatal mortality.We summarized this assessment in a 'Summary of Findings' table, created by the GRADEpro software. We rated the quality of the body of evidence for each selected outcome as high, moderate or low, or very low. Randomized studies were initially rated as high quality evidence but they were downgraded according to the five criteria listed below. Quasi‐experimental studies initially were rated as low quality evidence, but they could be upgraded if they did not have any serious methodological limitations. They could also be downgraded further.

There were five criteria that downgraded evidence (Atkins & GRADE, 2004):Risk of bias in individual studiesIndirectness of evidenceUnexplained heterogeneity or inconsistency of resultsImprecision of resultsHigh probably of publication bias

There were three criteria that upgraded the evidence for quasi‐experimental studies with no serious methodological limitations. (Atkins & GRADE, 2004):Large magnitude of effectPresence of a dose response relationshipsEffect of plausible residual confounding

Quality ratings, as determined by GRADE, are found in Table 4.

Subgroup analysis and investigation of heterogeneity

Heterogeneity was assessed based on clinical knowledge and theory and investigation of statistical criteria such as Tau², I² and significance of the Chi‐sqaure test.

Depending on data availability (> or equal to three studies per subgroup of interest), we conducted sub‐group analyses on the primary outcomes for the following variables:Age (10‐14 years, 15‐19 years, 20‐29 years, 30‐39 years, 40+)Geographical region (based on WHO regions)Sex of infantBaseline nutritional status–Anemic versus non‐anemic–Undernutrition versus normal nutrition, based on body mass index (BMI; BMI <18.5)–Low stature versus normal statureDuration of intervention–Women recruited prior to conception versus first trimester versus second trimester versus third trimester of pregnancyFrequency of intervention–Daily versus intermittent IFA supplementationDose of intervention–30 mg versus 60 mg elemental iron for IFA, MMN, or LNS supplementationUNIMMAP versus adapted UNIMMAP versus non‐UNIMMAP formulations for MMN supplementation (UNICEF, WHO, and UNU, 1999)–MMN supplements that contained a similar number and type of vitamins and minerals as the UNIMMAP formulation were categorized as 'adapted UNIMMAP' (+/−2 micronutrients, when compared to the standard UNIMMAP formulation)–Supplements with the same composition as UNIMMAP but different doses of vitamins and minerals were categorized as 'adapted UNIMMAP'

Variables were selected a priori, based on evidence to support their potential to impact the intervention effect. We carefully interpreted results from subgroup analyses. We also used meta‐regression techniques to assess how characteristics of studies (explanatory variables) may influence the size of the effect estimate (outcome variable). Potential variables may include the setting, dosing frequency, dosing form, compound, duration, route, sex of infant, SES status, or nutritional status.

Any subgroup analysis that was conducted post hoc was exploratory in nature and was stated as such.

Sensitivity analyses were conducted to determine whether the removal of studies with high risk of bias or the removal of non‐randomised studies significantly impacted findings. We defined studies as having a high risk of bias if one or more domains have been judged as 'high risk' or two or more domains have been judged as 'unclear risk'.

We did not include qualitative research.

We identified of 27,987 records from our database search and 670 records through hand‐searching and searching grey literature (Figure 3). Following title and abstract screening, full text screening was completed for 1,678 papers (1,246 from the database search and 432 from the grey literature and hand‐search combined). Of these, a total of 72 studies, with 314 associated papers, were identified for inclusion based on our pre‐defined inclusion and exclusion criteria. Eight studies were included in the review but did not contribute data to the meta‐analyses for various reasons: data was in a form that could not be pooled with other studies, no reported outcomes of interest, or insufficient number of studies to pool for meta‐analysis; at minimum, 3 studies were required to conduct a meta‐analysis for a given outcome (Diogenes et al., 2013; Duggan et al., 2014; Gowachirapant et al., 2017; Hambidge et al., 2019; Jarjou et al., 2006; Korkmaz et al., 2014; Prawirohartono et al., 2011; Taherian et al., 2002).

We identified a total of 72 studies (number of included papers = 314), involving 451,723 women as eligible for inclusion in this review (Figure 3). Eight trials were included in the review, but did not contribute data to the meta‐analyses (Diogenes et al., 2013; Duggan et al., 2014; Gowachirapant et al., 2017; Hambidge et al., 2019; Jarjou et al., 2006; Korkmaz et al., 2014; Prawirohartono et al., 2011; Taherian et al., 2002). Of these, Gowachirapant et al. (2017) was the only study that examined iodine supplementation versus placebo and Hambidge et al. (2019) was the only study that reported the effects of LNS versus placebo. Diogenes et al. (2013) and Taherian et al. (2002) were excluded from meta‐analyses because they were the only two studies that fit our eligibility criteria and evaluated the impact of calcium plus vitamin D versus placebo supplementation. Jarjou et al. (2006) examined calcium supplementation versus placebo (for which we have conducted meta‐analyses); however, this study reported outcomes that could not be pooled with other studies. Similarly, Korkmaz et al. (2014) was excluded from the iron versus placebo comparison and Prawirohartono et al. (2011) was excluded from the vitamin A versus placebo comparison and the zinc versus placebo comparison due to no common outcomes of interest. Duggan et al. (2014) was the only study that reported effects of vitamin B12 supplementation compared to placebo.

Twelve studies were conducted in the East Asia Pacific region (Dijkhuizen et al., 2001; Gowachirapant et al., 2017; Hanieh et al., 2013; Huy et al., 2009; Liu et al., 2013; Muslimatun et al., 2001; Prawirohartono et al., 2011; Supplementation with Multiple Micronutrients Intervention Trial (SUMMIT) Study Group, 2008; Sunawang et al., 2009; Tanumihardjo, 2002; Zeng et al., 2008; Zhao et al., 2015). Of these, 2 were excluded from analysis (Gowachirapant et al., 2017; Prawirohartono et al., 2011). Countries represented include: China (Liu et al., 2013; Zeng et al., 2008; Zhao et al., 2015), Indonesia (Dijkhuizen et al., 2001; Muslimatun et al., 2001; Prawirohartono et al., 2011; SUMMIT Study Group, 2008; Sunawang et al., 2009; Tanumihardjo, 2002), Vietnam (Hanieh et al., 2013; Huy et al., 2009), and Thailand (Gowachirapant et al., 2017).

Only 1 study was conducted in Europe & Central Asia, in Turkey (Korkmaz et al., 2014).

Seven studies were conducted in Latin America & the Caribbean (Belizán et al., 1997; Castillo‐Durán et al., 2001; Caulfield et al.,1999; Diogenes et al., 2013; López‐Jaramillo et al., 1997; Merialdi et al., 2004; Ramakrishnan et al., 2003). Of these, 1 was excluded from analysis (Diogenes et al., 2013). Countries represented include: Argentina (Belizán et al., 1997), Chile (Castillo‐Durán et al., 2001), Peru (Caulfield et al., 1999; Merialdi et al., 2004), Brazil (Diogenes et al., 2013), Ecuador (López‐Jaramillo et al., 1997) and Mexico (Ramakrishnan et al., 2003).

Thirteen studies were conducted in the Middle‐East and North Africa region (Aminisani et al., 2009; Asemi et al., 2013; Asemi et al., 2016; Falahi, 2010; Mohammad‐Alizadeh‐Charandabi et al., 2015; Naghshineh & Sheikhaliyan, 2016; Ouladsahebmadarek et al., 2011; Sabet et al., 2012; Sorouri et al., 2016; Taherian et al., 2002; Vaziri et al., 2016; Ziaei et al., 2007, 2008). Of these, 1 was excluded from analysis (Taherian et al., 2002). All 13 studies were conducted in Iran.

Nineteen studies were conducted in South Asia (Ahmad et al., 2016; Bhutta et al., 2009; Choudhury et al., 2012; Christian et al., 2003; Duggan et al., 2014; Hafeez et al., 2005a, b; Hossain et al., 2012; Khan et al., 2016; Kumar et al., 2009; Osendarp et al., 2000; Osrin et al., 2005; Roth et al., 2013 (AViDD); Roth et al., 2018 (MDIG); Sablok et al., 2015; Sahu et al., 2009; Tofail et al., 2008; West et al., 1999; West et al., 2011; West et al., 2014). Approximately half of the studies (n=8) were conducted in Bangladesh (Ahmad et al., 2016; Choudhury et al., 2012; Osendarp et al., 2000; Roth et al., 2013 (AViDD); Roth et al., 2018 (MDIG); Tofail et al., 2008; West et al., 2011; West et al., 2014). Four studies were conducted in India (Duggan et al., 2014; Kumar et al., 2009; Sablok et al., 2015; Sahu et al., 2009), 3 in Nepal (Christian et al., 2003; Osrin et al., 2005; West et al., 1999), and 4 in Pakistan (Bhutta et al., 2009; Hafeez et al.,2005a, 2005b; Hossain et al., 2012; Khan et al., 2016).

Eighteen studies were conducted in Sub‐Saharan Africa (Ashorn et al., 2015; Cox et al., 2005; Dewey, 2009; Darling et al., 2017; Etheredge et al., 2015; Fawzi et al., 2007; Friis et al., 2004; Huybregts et al., 2009; Jarjou et al., 2006; Kæstel et al., 2005; Kirkwood et al., 2010; Menendez et al., 1995; Moore et al., 2012; Preziosi et al., 1997; Roberfroid et al., 2008; Saaka et al., 2009; Semba et al., 2001; Zagré et al., 2007). Of these, Jarjou et al., 2006 was the only study excluded from analysis. Countries represented include Burkina Faso (Huybregts et al., 2009; Roberfroid et al., 2008), The Gambia (Jarjou et al., 2006; Menendez et al., 1995; Moore et al., 2012), Ghana (Cox et al., 2005; Dewey, 2009; Kirkwood et al., 2010; Saaka et al., 2009), Guinea‐Bissau (Kæstel et al., 2005), Malawi (Ashorn et al., 2015; Semba et al., 2001), Niger (Preziosi et al., 1997; Zagré et al., 2007), Tanzania (Darling et al., 2017; Etheredge et al., 2015; Fawzi et al., 2007) and Zimbabwe (Friis et al., 2004).

Two studies, Villar et al. (2006) and Hambidge et al. (2019) were multi‐country studies. Villar et al. (2006) was conducted in India, Peru, South Africa and Vietnam; Hambidge et al. (2019) was conducted in Democratic Republic of Congo, Guatemala, India and Pakistan.

A total of 439,649 women participated in the remaining 64 studies (Ahmad et al., 2016; Aminisani et al., 2009; Asemi et al., 2013; Asemi et al., 2016; Ashorn et al., 2015; Belizán et al., 1997; Bhutta et al., 2009; Castillo‐Durán et al., 2001; Caulfield et al., 1999; Choudhury et al., 2012; Christian et al., 2003; Cox et al., 2005; Mohammad‐Alizadeh‐Charandabi et al., 2015; Dewey, 2009; Dijkhuizen et al., 2001; Etheredge et al., 2015; Falahi, 2010; Fawzi et al., 2007; Friis et al., 2004; Hafeez et al., 2005a, 2005b; Hanieh et al., 2013; Hossain et al., 2012; Huy et al., 2009; Huybregts et al., 2009; Kæstel et al., 2005; Khan et al., 2016; Kirkwood et al., 2010; Kumar et alk., 2009; Liu et al., 2013; López‐Jaramillo et al., 1997; Menendez et al., 1995; Merialdi et al., 2004; Moore et al., 2012; Muslimatun et al., 2001; Naghshineh & Sheikhaliyan, 2016; Osendarp et al., 2000; Osrin et al., 2005; Ouladsahebmadarek et al., 2011; Preziosi et al., 1997; Ramakrishnan et al., 2003; Roberfroid et al., 2008; Roth et al., 2013 (AViDD); Roth et al., 2018 (MDIG); Saaka et al., 2009; Sabet et al., 2012; Sablok et al., 2015; Sahu et al., 2009; Semba et al., 2001; Sorouri et al., 2016; SUMMIT Study Group, 2008; Sunawang et al., 2009; Tanumihardjo, 2002; Tofail et al., 2008; Vaziri et al., 2016; Villar et al., 2006; West et al., 1999; West et al., 2011; West et al., 2014; Zagré et al., 2007; Zeng et al., 2008; Zhao et al., 2015; Ziaei et al., 2007, 2008).

Most of the outcomes were defined in the same way across different studies. The exception to this was in Christian et al., (2003) and Liu et al., 2013, where perinatal mortality was defined as stillbirths (from 28 weeks of gestation to delivery) and early neonatal death from birth to 6 days after delivery; this is compared to all other studies that defined perinatal mortality as stillbirths plus early neonatal deaths from birth to 7 days after delivery. For iron deficiency (measured by serum ferritin levels), Zhao et al., 2015 defined iron deficiency as a serum ferritin of <15ug/L, while the remaining studies (Falahi, 2010; Liu et al., 2013; Preziosi et al., 1997) reported it as a serum ferritin of <12ug/L. Lastly, Villar 2006 combined pre‐eclampsia and eclampsia cases; the other studies only considered pre‐eclampsia cases.

All studies reported sources of funding except: Asemi et al., 2013; Caulfield et al.,1999; Dijkhuizen et al., 2001; Falahi, 2010; Huy et al., 2009; Huybregts et al., 2009; Jarjou et al., 2006; Korkmaz et al., 2014; López‐Jaramillo et al., 1997; Menendez et al., 1995; Merialdi et al., 2004; Mohammad‐Alizadeh‐Charandabi et al., 2015; Muslimatun et al., 2001; Osendarp et al., 2000; Ouladsahebmadarek et al., 2011; Preziosi et al.,1997; Saaka et al., 2009; Sablok et al., 2015; Tanumihardjo, 2002; Villar et al., 2006; Ziaei et al., 2007, 2008.

Forty‐two studies included a statement of disclosure regarding potential conflicts of interest related to the study (Ahmad et al., 2016; Asemi 2016; Belizán et al., 1997; Castillo‐Durán et al., 2001; Christian et al., 2003; Dijkhuizen et al., 2001; Diogenes et al., 2013; Duggan et al., 2014; Etheredge et al., 2015; Fawzi et al., 2007; Friis et al., 2004; Gowachirapant et al., 2017; Hambidge et al., 2019; Hanieh et al., 2013; Hossain et al., 2012; Jarjou et al., 2006; Kæstel et al., 2005; Khan et al., 2016; Kirkwood et al., 2010; Korkmaz et al., 2014; Liu et al., 2013; Mohammad‐Alizadeh‐Charandabi et al., 2015; Moore et al., 2012; Ramakrishnan et al., 2003; Roberfroid et al., 2008; Roth et al., 2013; Hafeez et al., 2005a, 2005b; (AViDD); Roth et al., 2018 (MDIG); Sabet et al., 2012; Sablok et al., 2015; Sahu et al., 2009; Sorouri et al., 2016; SUMMIT Study Group, 2008; Tofail et al., 2008; Vaziri et al., 2016; West et al., 2011; Zhao et al., 2015; Ashorn et al., 2015; Dewey, 2009; Osrin et al., 2005; West et al., 2014; Zeng et al., 2008).

Of these, 37 studies (Ahmad et al., 2016; Asemi et al., 2016; Belizán et al.,1997; Castillo‐Durán et al., 2001; Christian et al., 2003; Dijkhuizen et al., 2001; Diogenes et al., 2013; Duggan et al., 2014; Etheredge et al., 2015; Fawzi et al., 2007; Friis et al., 2004; Gowachirapant et al., 2017; Hafeez et al., 2005a, 2005b; Hambidge et al., 2019; Hanieh et al., 2013; Hossain et al., 2012; Jarjou et al., 2006; Kæstel et al., 2005; Khan et al., 2016; Kirkwood et al., 2010; Korkmaz et al., 2014; Liu et al., 2013; Mohammad‐Alizadeh‐Charandabi et al., 2015; Moore et al., 2012; Ramakrishnan et al., 2003; Roberfroid et al., 2008; Roth et al., 2013 (AViDD); Roth et al., 2018 (MDIG); Sabet et al., 2012; Sablok et al., 2015; Sahu et al., 2009; Sorouri et al., 2016; SUMMIT Study Group, 2008; Tofail et al., 2008; Vaziri et al., 2016; West et al., 2011; Zhao et al., 2015) declared no conflict of interest amongst all authors. The five remaining studies (Ashorn et al., 2015; Dewey, 2009; Osrin et al., 2005; West et al., 2014; Zeng et al., 2008) indicated that one or several of the study's authors had a conflict of interest.

The remaining 30 studies (Aminisani et al., 2009; Asemi et al., 2013; Bhutta et al., 2009; Caulfield et al., 1999; Choudhury et al., 2012; Cox et al., 2005; Darling et al., 2017; Falahi, 2010; Huy et al., 2009; Huybregts et al., 2009; Kumar et al., 2009; López‐Jaramillo et al., 1997; Menendez et al., 1995; Merialdi et al., 2004; Muslimatun et al., 2001; Naghshineh & Sheikhaliyan, 2016; Osendarp et al., 2000; Ouladsahebmadarek et al., 2011; Prawirohartono et al., 2011; Preziosi et al., 1997; Saaka et al., 2009; Semba et al., 2001; Sunawang et al., 2009; Taherian et al., 2002; Tanumihardjo, 2002; Villar et al., 2006; West et al., 1999; Zagré et al., 2007; Ziaei et al., 2007, 2008) did not provide any statement of disclosure, and thus we could not comment on any potential conflicts of interest.