Study setting

Nepal has a growing population of about 28.5 million people and a gross per capita income of US $730.9 Makwanpur District is a rural district in the central region of Nepal covering 19 000 km² of mixed mountainous terrain, with a population of more than 500 000 people. It has one district hospital with a ratio of 7852 people per hospital bed. In 2001–2003 when the study was recruited, the neonatal mortality rate was 39/1000 live births, 94% of babies in rural areas were born at home and just 13% with a skilled birth attendant.5 Most households in Makwanpur District are dependent on subsistence agriculture.

Background to original trial

This is a longitudinal study of a cohort of 6436 maternal infant dyads recruited to a cluster Randomised Control Trial (RCT) of perinatal women's participatory groups in Makwanpur, Nepal, which showed a sustainable and low cost decrease in neonatal and maternal mortality.5 Detailed methodology of the original study is reported elsewhere.5 The study clusters were village development committees, local government divisions, of which there are 43 in Makwanpur District. Clusters were paired on basis of similar topography, ethnicity and population density. One cluster was excluded due to security concerns and 12 of the remaining 21 pairs were selected by random to the study. Each cluster had an average population of 7000 and area of 60 km². Clusters were mapped on foot between 1999 and 2000 to identify every household, followed by a baseline census to identify all female residents. In 2001, every eligible woman (married and aged 15–49) was interviewed and consented into the study. Recruitment took place between October 2001 and November 2003. Participants were visited monthly, and a pregnancy was registered when a woman ceased menstruation for 3 months in the absence of other explanation. Participants were interviewed 4 weeks postpartum for main outcomes of neonatal and maternal mortality. Thus, the baseline cohort was an entirely representative sample. In 2014, a long-term follow-up study of surviving and consenting participants was conducted and baseline (postpartum interview) and long-term follow-up data paired.

Long-term follow-up study

Two rounds of follow-up interviews took place across the study site, 24 pairs of interviewers visited families and if the family had moved attempts to trace them via neighbours were made. Seven supervisors directly observed 18% of all interviews. Data described here were collected during the first round of follow-up interviews and included anthropometry, forward and backwards digit recall testing, and a disability screening questionnaire.10 Interviewers were high school graduates and received a week's training in use of the screening tool.

Ethical approval for the study was obtained both from Nepal Health Research Council and the University College London research ethics committee. Informed verbal consent was granted by all participants prior to participation, which is more culturally appropriate in this setting than written consent.

Outcome

The main outcome was disability as measured at face-to-face interview using the Module on Child Functioning and Disability (MCFD).10 The MCFD is a new tool developed by the Washington Group on Disability Statistics and UNICEF, to improve and standardise information gathering on childhood disability. It is designed for use in children and young people aged 2–17 years, building on the Short Set of Questions that has been validated for use in adults and previously used in Nepal.11 We used it with permission from the Washington Group during the final stages of validation testing. It is a questionnaire taken by the child's main caregiver to assess functioning across a number of domains: speech and language, hearing, vision, cognition, mobility, self-care and emotions and behaviour. Owing to the current absence of validity data on the questions around emotions and behaviour, the UNICEF and the Washington Group have advised the definition of disability to be restricted to the core functional domains of speech and language, hearing, vision, learning, mobility and motor skills (MCFD core). Responses are usually ranked: no difficulty (1), some difficulty (2), a lot of difficulty (3), cannot do at all (4). We defined a positive disability screen as at least some difficulty (>=2), in at least one domain. This is to use the most inclusive definition of disability and because of the very low prevalence of other definitions in this study (supplementary appendix table A1)

Prenatal and perinatal risk factors

To minimise the pitfalls of post hoc testing, prior to reviewing the dataset a list of factors known to be associated with poor developmental outcomes was drawn up. Maternal characteristics, antenatal care and complications, problems during labour, essential newborn care, harmful traditional practices and neonatal illness were included.13 14 The dataset was then reviewed to see whether information on these factors was collected or if there were biologically plausible proxies available. Potential prenatal and perinatal risk factors were categorised into four groups: maternal, antenatal, labour and neonatal. Table 1 summarises the potential data available for each category and the final variables selected.

Of note, as 93% of infants were classified as ‘smaller than usual’ at birth, we chose ‘very small’ as the best marker for LBW. Gestational age (GA) was calculated based on last monthly period (LMP) as recalled by the mother, it is only available as completed month's gestation and 96% of infants were born with a reported GA of 9 months. For epidemiological-based studies in these types of settings, this is the only method available with which to calculate GA. Antenatal ultrasound was, and by and large is still, not available in this and similar settings and as the majority of babies were born at home, clinical gestational assessment at birth was also not an option. Very few infants were born at <8 months (0.7%) and for analysis all infants <9 months (<36 weeks) are classified as preterm. Skilled birth attendance is defined by WHO criteria and includes ‘mid-level’ workers such as health assistants but excludes community health workers who may nonetheless have undergone some obstetric training.15

Potential confounders

Potential confounders considered were trial allocation and baseline socioeconomic status (SES) as measured by household asset ownership score. Asset-based scores have been shown to be a reliable and simple way to assess differences in SES in LIC.16 17 Asset ownership was measured on a scale of 0–3, with1 indicating no assets on list, 1: basic assets such as clock or radio, 2: more expensive assets such as sewing machine or hand tractor and 3: very expensive assets such as a motor vehicle.

Anthropometric measures

Weight was measured to the closest 100 g using a Tanita HS302 Solar Scale with participants in minimal clothing, stood in the centre of the scale with arms at sides and face forward. Standing heights of participants were measured to the nearest mm against a Shorr board. Both height and weight are the average of two measurements unless they differed by more than 4 mm or 400 g, respectively, in which case a third was taken and the average of the two closest was used. Stunting was defined as a height for age z-score <−2 according to WHO 2006 standards. Weight at 1 month was standardised to create a mean of 0 and SD of 1.

Statistical analysis

Univariate analysis was performed by Pearson χ² for categorical variables, by two sample t-test for normally distributed continuous variables, and Mann-Whitney U test for non-normally distributed continuous variables. Collinearity was assessed by variance inflation factor scores with a score >10 indicating potentially significant collinearity. For some of the delivery variables such as hand washing, there were some ‘don't knows’ which are treated as missing data.

To build the final model, all the variables were ranked by their significance on univariate analysis and variables with a significance of <0.1 included in initial model. A backwards stepwise regression modelling approach was taken after included variables were checked for collinearity. Starting with the least significant variable on univariate analysis, variables were dropped and the model compared with the previous model by likelihood ratio testing. Variables were dropped permanently if the significance of the likelihood ratio test was ≥0.1 and retained if <0.1. The model was checked for confounding by trial allocation and SES. A variable was considered significantly confounded if the adjusted OR changed by more than 10% from the unadjusted model. The final models were model 1 (unadjusted), model 2 (adjusted by original trial allocation) and model 3 (adjusted by household asset score). Due to missing data, only 3318 of the 4419 children screened for disability were included in the regression model which was carried out as a complete case analysis.

To calculate population attributable risk (PAR), the relative risk (RR) was first estimated using log binomial regression. PAR (%) was then calculated using the formula [Pe(RR−1)]/[1+Pe(RR−10], where Pe is the prevalence of the exposure in the population that was screened for disability.

Statistical analysis was performed using Stata V. 13.18

Role of the funding source

The design of this study and all data collection, analyses, interpretation and the writing of the report were performed without the sponsors' involvement. Full access to data was granted to the corresponding author. All authors participated in the study design or analysis, and approved the submission of the manuscript.