Outcomes

The primary outcome of the trial is neonatal mortality, which is defined as the death of a liveborn baby during the first 28 completed days of age (population: C3, with a per-protocol analysis carried out in C4). See Table 4 for summaries of counting 28-week pregnancies, Table 5 for pregnancy outcomes including neonatal mortality, and Table 6 for subgroup analyses (see the “Subgroup analyses” section). Table 7 shows pregnancy outcome by gender.

Table 4 Pregnancies (number (%) unless otherwise stated)Table 5 Pregnancy outcomes, perinatal mortality, and neonatal mortality (number (%) unless otherwise stated)Table 6 Neonatal mortality rate (deaths per 1000 live births) by subgroup, with interaction tests (population C3)Table 7 Details of counting children (population C1) (number (%) unless otherwise stated)

The following secondary outcomes are to be formally tested and a 95% confidence interval constructed:

Maternal mortality [12], defined as death of a woman whilst pregnant or within 42 days of termination of pregnancy, irrespective of the duration and site of pregnancy (e.g. fallopian tube, uterus), from any cause related to or aggravated by the pregnancy or its management but not from accidental or incidental causes (see Table 8; population: P1);

Stillbirths and perinatal deaths [12], with a stillbirth defined as the death of a baby after 28 weeks of pregnancy but before or during birth, and a perinatal death defined as either a stillbirth or the death of neonate before 7 completed days of age (see Table 5; population: C1);

Hospital admissions (other than for delivery) of enrolled women (during pregnancy and in the immediate post-natal period) and their babies (in the immediate post-natal period) (these are events defined a priori as serious adverse events; see Table 9; populations: P1 for women and C1 for babies);

Maternal blood transfusion during pregnancy and in the immediate post-natal period (this is one of the events defined a priori as serious adverse events; see Table 9; population: P1);

Table 8 Maternal mortality (population P1) (number (%) unless otherwise stated)Table 9 Serious adverse events (number (%) unless otherwise stated)

The following secondary outcomes are to be tabulated but not formally tested:

Neonatal causes of death (see Table 5; population: C1) and maternal causes of death (see Table 8; population: P1);

Antenatal and postnatal care of mother (e.g. number of antenatal care visits, antenatal care provided; care provider; see Table 10; population: P1);

Delivery care (e.g. use of skilled birth attendant, place of delivery, clean delivery practices; see Table 11; population: C1);

Postnatal care of neonate (e.g. umbilical cord care, thermal care, breastfeeding, care seeking; see Table 11; population: C1);

Health knowledge (see Table 12; population: P1);

Cost effectiveness of the intervention.

Summaries of mothers’ antenatal and postnatal care are shown in Table 10. Summaries of delivery, babies’ immediate newborn care, and babies’ care in the first month are shown in Table 11. The data in these two tables were collected for all cases of neonatal death and stillbirths, all multiple births, and a random sample of 10% of other live neonates.

Table 10 Mothers’ antenatal and postnatal care (P1) (number (%) unless otherwise stated)Table 11 Delivery, immediate newborn care, and newborn care during baby’s first month (C1) (number (%) unless otherwise stated)Table 12 Mothers’ knowledge and attitudes (P1) (number (%) unless otherwise stated)Analysis methods

Neonatal mortality in this trial has a complex four-level hierarchical structure, with multiple women per cluster, potentially multiple pregnancies per woman, and potentially multiple births per pregnancy. The effect of the active intervention on neonatal mortality (compared to usual care) will be estimated using a generalised estimating equations (GEE) analysis approach. This allows for non-independence of outcomes from the same cluster and for non-independence of multiple outcomes from the same woman. Mixed models (with cluster as a random effect, which are also termed hierarchical or multilevel models) are perhaps more commonly used than GEEs for the analysis of cluster randomised trials. The advantage of the GEE/robust standard error approach here is that the four-level hierarchical structure does not have to be explicitly modelled, so avoiding potential convergence problems.

In detail, the relative risk with a 95% confidence interval will be obtained from a GEE model with a binary outcome, a log link, a “working” assumption of independence with robust standard errors to take account of clustering at village level. The model will include the stratifying variables, which were village size and distance to the nearest community health centre or civil hospital [13, 14].

Secondary analyses will extend the GEE model for the primary outcome described above to (separately) investigate interactions by the randomisation stratifiers, whether women were enrolled pre- or post-randomisation, gender, caste, wealth, and male and female primary caregiver literacy (see below).

The risk difference with a 95% confidence interval will be obtained from a GEE model with a binary outcome, an identity link, a “working” assumption of independence with robust standard errors to take account of clustering. This model will also include the stratifying variables.

Secondary outcomes that are binary will be analysed using the same approach as for the primary outcome.

Maternal mortality in each arm will be expressed as the number of maternal deaths per 100,000 live births, with the ratio of these computed. A nonparametric bootstrap confidence interval (bias corrected and accelerated, 2000 replications at cluster level, stratified by randomisation arm) will be constructed for the ratio (on a log-transformed scale) of these arm-specific mortality rates.

Adjustment for covariates

All comparisons between trial arms will adjust for the stratification factors (village size and distance to the nearest community health centre or civil hospital (both binary)) and no others.

Methods used for assumptions to be checked for statistical methods

The models used for the continuous outcomes assume that residuals are normally distributed. Robust standard errors allow for potential heteroscedasticity according to levels of predictor variables but do make an assumption of normality conditional on levels of predictor variables. This assumption will be checked by examination of appropriate quantile-quantile plots of standardised residuals. The central limit theorem ensures that results are robust provided that violations of the normality assumptions are not substantial. Minor violations, even if statistically significant, are of little practical consequence. For this reason, formal hypothesis tests of normality assumptions will not be carried out.

Alternative methods to be used if distributional assumptions do not hold

Nonparametric bootstrap confidence intervals (bias corrected and accelerated, 2000 replications at cluster level, stratified by randomisation arm) will be reported if the normality assumptions are seriously violated.

Sensitivity analyses for each outcome where applicable

In the primary analysis, missing data will not be imputed. In secondary analyses of the primary outcome and key secondary outcomes, multiple imputation by chained equations (MICE) will be used if missingness is greater than 5%, as has been recommended [15]. For analysis of clustered data, it is important that the model for imputation includes cluster-specific random effects [16]. Such analyses will be carried out using the Jomo package within the statistical software environment R [17]. Imputation will be carried out separately in each trial arm. Auxiliary variables to potentially be used will include the randomisation stratification factors, caste, gender, male and female primary caregiver literacy and education, the wealth indices, the adherence to intervention variables defined above, being a twin, and previous miscarriage/termination/stillbirth/neonatal death.

If the effect of the intervention is statistically significant, and remains so in the MICE analysis detailed above, then the multiple imputation analysis will also be extended to determine the amount of bias over and above that allowed for by the multiple imputation model that would render the primary analysis non- statistically significant.

Subgroup analyses

We will conduct exploratory subgroup analyses of the primary outcome by:

Village population (binary, as used in the stratified randomisation),

Distance to nearest community health centre/civil hospital (binary, as used in the stratified randomisation),

Whether women were enrolled pre- or post-randomisation,

Gender,

Caste,

Wealth index 1 (in three categories determined by the material the house is made of),

Wealth index 2 (in five categories determined by the number of relevant items owned by the household, with the interaction tested using a trend test).

Primary female caregiver literacy in 3 groups. This to be replaced by female education if more than 10% of the participants have a missing value for literacy and education status is not missing,

Primary male caregiver literacy in 3 groups. This to be replaced by male education if more than 10% of the participants have a missing value for literacy and education status is not missing.

For each of the above factors, statistical tests for interaction will be carried out, with claims of different effects in subgroups only made if there is strong evidence (p < 0.01) of an interaction. See Table 6.

Additional analyses

The risk difference (and its 95% confidence interval) will be multiplied by the number of live births in the intervention arm to give an estimate (and 95% confidence interval) for the number of lives saved.

Additional analyses will include an economic evaluation. A cost-effectiveness calculation in terms of cost per neonatal death averted and cost per life year saved will be conducted (cost per disability-adjusted life year saved will not be considered as no measure of future disability is available). The sensitivity of these outcomes to the most important inputs—labour costs and exchange rate movements—will be examined.

The direct additional provider costs of the CHAMPION intervention activities compared to existing standard of care in the control arm will be calculated. Total spending will be cross-checked with funding sources for accuracy. Equipment and other resources provided to clinics which benefitted both control and intervention villages will be noted separately. Spending will be divided into running costs and capital costs. Start-up costs are limited and are assumed to be fully depreciated during the trial because the NICE Foundation had previously implemented a similar version of the programme elsewhere in India. Straight line depreciation of capital equipment (computers, ambulances, and medical equipment will be based on 3-, 4-, and 8-year lifespans, respectively) will be allowed for, consistent with usual account practices. Capital spending outside these items is assumed to be fully depreciated immediately. There are no contributions in kind.

Annual cost figures will be adjusted by India’s GDP deflator in order to convert values to July 2023 rupees. Average exchange rates from July 2023 will be used to convert rupee figures to US dollars.

Statistical software

Stata version 18 (StataCorp. 2023. Stata Statistical Software: Release 18. College Station, TX: StataCorp LLC) and/or R (R Core Team 2022. R: A language and environment statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.)