Study design and sampling strategy

This study was led by the Global Alliance for Improved Nutrition with partner organisations St John’s Research Institute, Bangalore, India (SJRI), The India Nutrition Initiative and Cornell University. The study consisted of a cross-sectional survey of non-pregnant mothers of reproductive age (MRA) 18–49 years of age and children 6–59 months of age (preschool-age children (PSC)) living in rural and urban Uttar Pradesh. A total of twenty-five districts were selected: ten districts were purposefully sampled as part of a baseline evaluation to a double fortified salt programme in the state^{(Reference Cyriac, Webb Girard and Ramakrishnan12)} (details forthcoming) and fifteen additional districts were selected randomly using simple random sampling with a random number generator out of all remaining districts in the state, for a total of twenty-five districts. Within each district, five villages (rural) and five wards (urban) were randomly selected from the 2011 Census of India. Within each ward, one Census Enumeration Block was selected randomly using the 2011 Census of India. Within each of these villages and Census Enumeration Blocks, five households were selected to be included in the survey using a modified random walk method (see online supplementary material, Supplemental Fig. 1).⁽¹³⁾ In nine wards, an insufficient number of households fit our inclusion criteria (i.e. there were fewer than five households containing a mother of reproductive age with at least one child 6–59 months of age). In these cases, where possible, we attempted to select additional households from neighbouring villages or wards selected for the survey in the same district. As such, the total number of participants included in the study varied slightly from the original design. Within each household, one non-pregnant mother of reproductive age and one of her PSC 6–59 months of age were included. Households were excluded only in the case of physical or mental impairment that could impede measurements or provision of informed consent. Data were collected between October and December 2016.

A sample size of 1250 households in total was calculated to obtain a 5 % precision on the prevalence of Fe deficiency, in twenty-five randomly selected districts, anticipated prevalence of 25 %^{(Reference Haas, Rahn and Venkatramanan14)}, average household size of 5·7⁽¹¹⁾ and a design effect of 2^{(Reference Pasricha, Drakesmith and Black15)}. Using this pre-specified sample size of 1250, for this analysis, we were able to detect a minimum correlation coefficient with Hb of 0·09, with 90 % power and 5 % precision.

Measurements and data collection

All MRA were interviewed with a household survey. The survey consisted of questions about household demographic characteristics such as socio-economic status, water, sanitation, hygiene, use of Integrated Child Development Services (a programme delivered through the Anganwadi Centres), use of other safety net programmes, as well as information on the mother and child’s medical history and use of nutritional supplements (i.e. Fe, folic acid, Zn, Ca, vitamin C, vitamin D and multivitamins). Women’s dietary diversity was assessed based on a 9-food group score for a 24-h recall period^{(Reference Kennedy, Ballard and Dop16)} and for children a 7-item child dietary diversity score was assessed according to WHO guidelines at the time during a 24-h recall period⁽¹⁷⁾. Food insecurity in the past 12 months was measured using an 8-item Food Insecurity Experience Scale⁽¹⁸⁾.

Plax-Cruzer Series (Indman Scale Corporation) weighing scales were used to assess weight of the mother and child to the nearest 0·1 kg, using the tare function for child weight. Standing height was taken from MRA and children 2 years and above using the Seca 213 stadiometer to the nearest 0·1 cm. Recumbent length of children less than 2 years of age was measured using the Seca 417 infantometer (Seca) to the nearest 0·1 cm. Stadiometers, length boards and weighing scales were calibrated daily using standard procedures provided by the manufacturers. Reliability estimates for anthropometric measurements were obtained by comparing the supervisors’ measurements (only supervisors performed anthropometry measurements) to those of an expert when assessing the same child, using five children^{(Reference de Onis, Onyango and Van den Broeck19,Reference Fernald, Kariger and Engle20)} . Reliability measurements for weight and height yielded a Pearson’s correlation coefficient between measurements of the data collectors and expert of > 0·98 and a coefficient of reliability of > 0·98.

Hb concentration was measured using the HemoCue 201 + Analyser (HemoCue). A non-fasting venous blood sample (10 mL) was collected from each MRA and 7 mL was collected from each PSC. First, 4 mL of venous blood was collected in plastic spray-coated K₂EDTA tubes from BD Vacutainer^®. After slowly inverting the K₂EDTA vial twice, a syringe was used to collect a few drops of whole blood from the tube and to place them on a clean plastic slide. The phlebotomist then filled a microcuvette with the drop and immediately tested the Hb concentration using the HemoCue. A drop of whole blood was also used to test for the presence of malaria, specifically P. vivax and P. falciparum, using a rapid diagnostic kit (FalciVax immune-chromatography, Zephyr Biomedicals) on both women and children.

An additional 3–6 mL of venous blood was collected in trace-element free tubes, with increased silica act clot activator, silicone-coated interior from BD Vacutainer®.

Laboratory blood measurements

Laboratory blood analyses were conducted at SJRI (i.e. red blood cell folate) and Sanjay Gandhi Postgraduate Institute (SGPGI) (i.e. all other analyses). Internal laboratory quality control was performed according to standard procedures. The laboratory at SGPGI completed the RIQAS certification through Randox Laboratories Limited (United Kingdom). The laboratory at SJRI participated in the Centers for Disease Control and Prevention’s VITAL-EQA programme.

Serum ferritin and vitamin B₁₂ were measured using sandwich electrochemiluminescence immunoassay ELECSYS 2010 (ROCHE Diagnostics). Serum C-reactive protein (CRP) and α1-acid glycoprotein (AGP) were measured using immunoturbidimetry: Cobas Integra 800 for CRP and Hitachi 902 for AGP (ROCHE Diagnostics). Serum Zn was measured using atomic absorption spectroscopy Hitachi Model Z2300 (Hitachi Group). β-thalassaemia was measured using ion-exchange HPLC VARIANT II β-thalassaemia Short Programme (BIO-RAD Laboratories). Serum retinol was measured using reverse-phase HPLC: Dionex Ultimate 3000 series HPLC instrument (Thermo Fisher Scientific). Red blood cell (RBC) folate was quantified using the WHO-recommended folate microbiological assay^{(Reference Cordero, Crider and Rogers21)}. The folate concentration was determined by measuring the turbidity of the inoculated medium at 590 nm using a microplate reader (SYNERGY H1 microplate reader, BioTek Instruments). The assay was carried out in a 96-well microplate and calibrated using 5-methyltetrahydrofolic acid standard provided by the CDC^{(Reference O’broin and Kelleher22,Reference Pfeiffer, Zhang and Lacher23)} .

Variable creation and statistical analyses

A Wealth Index, using five categories, was calculated using principal component analysis with family assets, house materials and number of rooms, and agricultural land ownership, home ownership or commercial ownership. A water, sanitation and hygiene (WASH) index, also using quintiles, was calculated using principal component analysis with improved or unimproved source of drinking water, toilet facility and hygienic appearance of the mother and her child and the household compound^{(Reference Webb, Stein and Ramakrishnan24)}. BMI for women was derived by dividing the weight of mothers (in kg) by the square of their height (in cm). The following BMI categories were used to classify women: < 18·5, 18·5–24·9, 25·0–29·9 and over > 30 kg/m². Length-for-age, weight-for-length and weight-for-age z-scores were calculated using the WHO 2006 child growth standards. Z-scores less than –2 were used to define stunting, wasting and underweight, respectively, and z-scores less than –3 were used to define severe stunting, severe wasting and severe underweight^{(Reference de Onis, Onyango and Van den Broeck19)}.

Anaemia was defined according to WHO criteria⁽²⁵⁾ in MRA: any anaemia as Hb < 12 g/dL, mild anaemia as Hb 11–11·9 g/dL, moderate anaemia as Hb 8–10·9 g/dL and severe anaemia as Hb < 8 g/dL; in PSC, any anaemia as < 11 g/dL, mild anaemia as Hb 10–10·9 g/dL, moderate anaemia as Hb 7–9·9 g/dL and severe anaemia as Hb < 7 g/dL. Hb adjustment for altitude was not needed as all areas were below 1000 m above sea level⁽²⁵⁾. Fe deficiency was defined as ferritin < 15 µg/L in MRA and < 12 µg/L in PSC⁽²⁶⁾. Vitamin A deficiency was defined as serum retinol concentration < 0·7 µmol/L in MRA and PSC⁽²⁷⁾. Ferritin (children and women) and retinol (children only) were adjusted for inflammation (CRP and AGP) using the BRINDA regression approach^{(Reference Suchdev, Namaste and Aaron28,Reference Namaste, Rohner and Huang29)} . Vitamin B₁₂ deficiency was defined as vitamin B₁₂ concentration < 150 pmol/L in MRA and PSC^{(Reference de Benoist30)}. Zn deficiency was defined as Zn concentration < 10·7 µmol/L in MRA and < 9·9 µmol/L in PSC^{(Reference Brown, Rivera and Bhutta31)}. Individual Zn values were adjusted to a non-fasting state and 20 min between collection and processing for both women and children^{(Reference McDonald, Suchdev and Krebs32)}. In children only, they were also adjusted for CRP and AGP using the BRINDA regression approach^{(Reference McDonald, Suchdev and Krebs32)}. RBC folate level to define folate deficiency (for all age groups) using macrocytic anaemia as haematological indicator was defined as < 305 nmol/L^{(Reference Pfeiffer, Sternberg and Hamner33)}. Elevated CRP was defined as CRP concentration > 5 mg/L⁽³⁴⁾. Elevated AGP was defined as AGP concentration > 1 g/L^{(Reference Erhardt, Estes and Pfeiffer35)}. β-thalassaemia was defined as Hb A2 variant > 3·2 %^{(Reference Bain36)}.

We compared the magnitude of the associations between various predictors by fitting the data to a pre-specified model to determine their association with one another and with Hb (see online supplementary material, Supplemental Fig. 2). The pre-specified model was based on a review of the literature, and the pathways leading to anaemia were proposed drawing on the UNICEF conceptual framework of malnutrition that highlights the immediate, underlying and basic causes of malnutrition⁽³⁷⁾. Explanations of the theory behind the model development are presented in online supplementary material, Supplemental Fig. 2.

Univariate and bivariate analyses were conducted using SAS 9.4 (SAS Institute), and the path model was developed in StataSE version 15 (StataCorp) using maximum likelihood estimation. First, means, standard deviations and prevalence statistics were derived using standard SAS procedures. Non-normally distributed blood biomarkers (ferritin, vitamin B₁₂, RBC folate, Zn, retinol, CRP, AGP) were natural logarithm (ln) transformed and geometric means were reported. Second, multiple variable logistic regression models were used to estimate the factors associated with anaemia in children and women, and population attributable fractions were calculated for each of these factors^{(Reference Bruzzi, Green and Byar38)}. Variables from our pre-specified theoretical model (see online supplementary material, Supplemental Fig. 2) were included in the multiple variable models. In all analyses, we accounted for clustering at the district level and weighted for district and setting (urban and rural) sampling. Weights were constructed for rural and urban populations separately using adjustment factors for (1) the proportion of people living in rural and urban localities in each district, (2) the size of each district and (3) the district’s likelihood of being selected. In SAS, proc surveymeans/surveyfreq and proc surveyreg/surveylogistic commands were used to account for the complex survey design; in StataSE, the svy command was used.

Third, the hypothesised path model examined the direct and indirect associations among WASH, dietary diversity, food insecurity, CRP, AGP, nutritional biomarkers (ferritin, Zn, retinol, vitamin B₁₂, RBC folate) and Hb concentration. Prior to running the path model, Pearson’s correlations were run to examine correlations between all variables included in the model. A single model for PSC and for MRA was fitted to the data. Model fit was evaluated with common standards: a comparative fit index^{(Reference Bentler39)} and a Tucker Lewis index > 0·90 for acceptable fit and > 0·95 for good fit^{(Reference Tucker and Lewis40)}, and a standardised root mean squared residual^{(Reference Lt and Bentler41)} and a root mean square error of approximation < 0·08 for acceptable fit and < 0·05 for good fit^{(Reference Steiger42)}. If model fit statistics were not acceptable, modification indices were examined to decide whether additional pathways should be examined to improve model fit, and the model was re-specified. The PSC model was adjusted for child age in months. Group comparison for rural and urban settings was used to determine whether the model may be different in the two populations. A Wald test was used to determine statistically significant (P-value < 0·05) differences for any particular path. Standardised and unstandardised estimates were calculated.

β-thalassaemia trait was measured in only a subset of the women (n 744) due to insufficient blood, and the prevalence of β-thalassaemia trait was low in this population (3 %); hence, a path analysis with β-thalassaemia was run separately. The model with thalassaemia included direct pathways from β-thalassaemia trait to Hb concentration and to ferritin concentration.

A further additional analysis was run using full information maximum likelihood. Several laboratory analyses were missing for some participants, particularly children, given insufficient blood sample. Therefore, 45 % of PSC and 19 % of MRA were missing at least one value for the predictors of interest. Additional path modelling was done in StataSE using the sem command and the mlmv option, which estimates coefficients using full information maximum likelihood with all available data under the assumption that any missing data were at random^{(Reference Enders and Bandalos43)}.