Survey design and setting
The Women’s Work in Agriculture and Nutrition (WWN) study is a longitudinal study of mother-infant dyads living in irrigated rural areas of Sindh province, Pakistan. Here we use data from the baseline cross-sectional survey, conducted from December 2015 to February 2016.
Sampling
A sample size of 1000 dyads was calculated to detect a difference in maternal BMI of 0.18 for every additional hour worked with 80% power at a 5% level of significance [4]. This sample size provides adequate power to explore factors associated with maternal and infant nutritional status and to perform statistical mediation analysis [13]. Further information on the sampling strategy is provided in Additional file 1.
Participants were selected via systematic random cluster sampling. Initially, administrative villages with perennial canal irrigation were selected; villages with a population below the 10th and above the 90th percentiles of estimated village sizes were excluded, and random sampling was used to select villages to provide the estimated sample size of 1000 mother-infant dyads. Villages with perennial canal irrigation were chosen as the study site because women in these villages are frequently involved in commercial agriculture, including cotton harvesting. All dyads in selected villages were invited to participate in the study if: (i) the infant was a singleton birth ≥2 weeks and ≤ 12 weeks of age on the day of the first interview; healthy without congenital deformations that would impact on their ability to eat and (ii) the primary caregiver (i.e. the biological mother) intended to reside in the study area over the next 10 months.
Questionnaire and spot observations
An interviewer-administered questionnaire was collected from mothers using electronic data capture (Samsung tab-4) to obtain information relating to: socio-demographic characteristics, household food insecurity, health, maternal 24-h food consumption and maternal agriculture work history during pregnancy (Additional file 2). The questionnaire took approximately 60 min to complete. Spot observations were also performed to record housing materials and the hygienic conditions of the environment.
Anthropometric assessment
Two serial measurements of maternal and infant weight and height/length were collected, following standard procedures [14], by trained fieldworkers who were selected based on their technical error of measurement (TEM) results following a 5-day training programme. After removing shoes and heavy clothing, maternal weight was measured to the nearest 0.5 kg using digital electronic scales (Tanita digital bathroom scale); maternal height was measured to the nearest 0.1 cm using a portable stadiometer (Seca 213); infant weight was measured to the nearest 0.01 kg using digital electronic scales (LAICA weight scale for babies) and infant length was measured to the nearest 0.1 cm using an infantometer (Seca 416). A third measurement was taken if the difference between the first two measurements was above a pre-defined threshold (i.e. > 0.7 cm for maternal height and infant length; > 0.5 kg for maternal weight; > 0.1 kg for infant weight); and an average of the two closest measurements was calculated. For women who refused to remove heavy jewellery and/or clothing (n = 134; 11.7%), their measured weight was reduced by 0.5 kg for heavy jewellery and 0.5 kg for heavy clothing. No adjustments for clothing were made to infant weight.
Data management and analyses
Anthropometric characteristics
Maternal BMI post-pregnancy was calculated (weight (kg)/(height (m2)); and women were classified as underweight, normal weight, overweight and obese using age-specific international cut-off points for adolescents aged less than 18 years [15, 16]; and the World Health Organization (WHO) adult cut-offs for women aged 18 years or above [17].
Infant LAZ and weight-for-length (WLZ) were generated using the WHO growth standards [14]. Data for infants with biologically implausible anthropometric results were excluded (i.e., those with z-scores <− 6 SD or > + 6 SD for LAZ and those with z-scores <− 5 SD or > + 5 SD for WLZ) [14]. Infants were classified as stunted or wasted if their z-scores for LAZ or WLZ were < − 2 SD, respectively [14].
Socio-demographic characteristics
A household wealth index was created using factor analysis applied to proxy indicators of the household environment (ownership of consumer durables; house ownership; land ownership; main source of energy for cooking; livestock ownership; electricity; source of drinking water and type of toilet facilities; number of rooms used for sleeping; type of materials used for floors, the roof and walls). Socio-economic status (SES) quintiles were created; and internal validity was checked by assessing ownership of durable assets and housing characteristics by SES quintile. Maternal education was not included in the creation of the household wealth index because of its known independent effect on nutrition and health outcomes.
Food insecurity and maternal dietary diversity
To measure food insecurity, answers to questions on anxiety and uncertainty about the household food supply and insufficient food quality experienced in the past 30 days were collected [18]. Binary variables (yes/no) were created to capture the proportion of households who ate a limited variety of foods due to a lack of resources in the past 30 days and to capture the proportion of households who worried about not having enough food in the past 30 days. The Household Food Insecurity Access Scale was not generated because we collected only a sub-set of the questions to reduce the length of the questionnaire and respondent burden.
To measure maternal dietary diversity, a single semi-qualitative 24 h dietary recall was used. The Minimum Dietary Diversity score (ordinal variable) was then generated for women of reproductive age, using reported intakes of foods and beverages during the past 24-h [19].
Agricultural work data
Binary variables (yes/no) were created for reported engagement during pregnancy in livestock-related agricultural activities, any crop-related agricultural activities (including cotton harvesting), and cotton harvesting alone. The livestock-related activities included fodder collection/preparation/chopping, animal washing, milking animals, providing care to animals, grazing animals, giving water to animals and egg collection. The crop-related activities included sowing, transplanting, digging, weeding, applying fertilizer, grain harvesting, vegetable harvesting and cotton harvesting. Cotton harvesting, which is almost exclusively done by women in this part of Pakistan, was examined separately because it involves particularly long hours of intensive work under the sun during the summer/autumn months (July to November).
Statistical analyses
Hypothesized models of pathways related to maternal BMI and infant size were drawn based on the available literature and discussion with experts (Additional file 3). These diagrams were drawn using the DAGitty software and constitute directed acyclic graphs (DAG). These DAGs provide a means of visually identifying common causes and effects in a set of variables within a multivariable model and are used to identify variables to be included in multivariable analysis to estimate the total effect of pre-specified exposures of interest [20].
Associations of agricultural-related work with maternal BMI and infant LAZ were examined using univariable and multivariable linear regressions. We restricted the analysis to LAZ (in contrast to WLZ) as LAZ was more likely to reflect maternal conditions (i.e. agricultural work) during pregnancy. WLZ is more susceptible to any adverse exposures in the immediate environment after birth and may mask the effects of maternal agricultural work in pregnancy.
The variables included in the multivariable models were retained based on the information provided by the DAGs (Additional file 3). Some pre-specified covariates were included in the model due to their known effect on maternal BMI (i.e. maternal age and number of days post-partum) and infant growth (i.e. infant age and sex, maternal height) [21]. We adjusted for the number of weeks post-partum (2–12 weeks) at which maternal weight was collected to address potential bias in associations with maternal BMI. The multivariable analyses were conducted on the complete case sample.
Structural-equation models were used to test whether maternal BMI (post pregnancy) mediated the association between agricultural work during pregnancy and infant LAZ at 2–12 weeks of age. The total effect was decomposed into direct and indirect effects. We ensured that the temporality of the variables in the model was respected. Maternal workload during pregnancy precedes both maternal BMI post-pregnancy (mediating factor) and infant LAZ at 2–12 weeks (outcome). Both maternal BMI and infant LAZ were measured around the same time and so it is difficult to prove that maternal BMI precedes infant LAZ but maternal BMI post-pregnancy was used as a proxy for maternal nutritional status during pregnancy. From a biological point of view, maternal BMI would predict infant LAZ rather than vice versa. SEM models were driven by both our initial hypothesis and by the significance of the initial associations.
Non-parametric bootstrapping was used to estimate confidence intervals (CIs) and corresponding p-values in the univariable and multivariable analyses [22]. Robust standard errors were used to account for clustering at the village level. All analyses were conducted using Stata/IC (version 14.1). The type I error risk was set at 0.05.