Adhering to the CONSORT-Equity 2017 reporting standards, which aims to improve the reporting of intervention effects in randomized trials where health equity is relevant, we collected our data in Faridabad and Palwal district of Haryana in India between 30 July 2015 and 31 October 2018 as part of the individually randomized controlled parallel-arm ciKMC trial [3, 11, 13]. Earlier studies done in this area show substantial health inequity across characteristics like caste, gender, maternal literacy, and wealth status [9, 13, 14]. Our formative research facilitated capturing the characteristics of the PROGRESS-Plus framework, including caste, gender, religion, education, and socioeconomic status [15, 16]. The PROGRESS-Plus framework is an acronym used to identify characteristics that stratify health opportunities and outcomes [16].
The main effect estimate of the current study was the difference in concentration index between infants in the ciKMC and control arms for post enrolment survival until 180 days of life. We also compared the post-enrolment mortality rate ratios until 180 days of life between the infants in the ciKMC and the control arms across wealth status, maternal literacy, family’s caste, family’s religion, and infant’s sex.
The field team assessed the infants at home and weighed them as soon as possible (no later than 72 h) after birth; they were eligible if they weighed between 1500 and 2250 g [11]. Infants with an inability to feed, difficulty in breathing, less than normal movements or with gross congenital malformation; those for whom KMC had been initiated in hospitals; and infants whose mothers planned to move out of the study area during the trial period were excluded.
The intervention consisted of the newborns being kept in skin-to-skin contact with their mother or a surrogate and exclusively breastfed for as long as possible. An intervention delivery team made nine home visits in the intervention arm during the first 28 days of life to support KMC. No intervention was given to the control families but families in both the intervention and control arms of the trial were expected to receive routine home-based care from the public health system, which comprises of 6 home visits on day 1, 3, 7, 14, 21 and 28 of life [17]. We collected socioeconomic and demographic data at baseline. During regular home visits, a separate team of well-trained research assistants, masked to trial-arm allocation, collected data on mortality of the participating babies until they were six months of age. The data collection procedures were identical in both arms.
Descriptive statistics with summary measures of health inequality
We used an asset index score, a composite measure of the living standards of the households, to rank the study participants. We calculated the asset index, using data on household ownership of selected assets (e.g., televisions and bicycles), the materials used for housing construction, sanitation facilities and the source of drinking water. Each household asset was assigned a weight or factor score generated through principal components analysis. The resulting asset scores were standardized to a standard normal distribution with a mean of zero and a standard deviation of one. These standardized scores were then used to divide the study population into five quintiles. The method we used to generate the asset index was similar to that used by the Demographic and Health Survey Program [18]. The wealth status of the lower 40% of the study population based on asset index score (i.e., representing the two lowest quintiles) was categorized as poor – the upper three quintiles were categorized as non-poor [19].
We present the study outcomes by wealth quintile to explore social gradients in the two trial arms. To investigate, summarize and draw inferences about the impact of the intervention on health inequity, we used concentration curves, concentration indices and the difference in the concentration indices between the two arms [20]. The concentration curve plots the cumulative proportion of the health variable (y-axis) against the cumulative percentage of the population, ranked by living standards, beginning with the poorest, and ending with the richest (x-axis) [21]. The concentration index is defined as twice the area between the concentration curve and the line of equality (the 45-degree line). So, when there is no socioeconomic-related inequality, the concentration index is zero [20, 21]. We used an F-test to estimate the statistical precision i.e., 95% confidence interval of this difference in concentration index (Δ ci) for mortality up to 180 days of life between the intervention and control arm. A positive Δ ci reflects a positive equity impact i.e., reduced inequity and negative Δ ci indicates increased inequity [22]. The magnitude of the Δ ci is a measure of the extent to which inequity was increased or decreased due to the intervention. We used Stata 16.1 (StataCorp LLC, College Station, Texas) and community-contributed packages (“DASP” and “Lorenz”) for our analyses [23, 24].
Inferential analysis
We frequency-aggregated the data for death, follow-up time and infants enrolled from the same household across the subgroups defined by wealth status [non-poor vs poor], family caste [scheduled caste (SC)/scheduled tribe (ST)/other backward caste (OBC) vs other], mother’s literacy [illiterate vs literate], infant’s sex, and religion [Hindu vs other]. In each stratum, we then estimated the incidence rate ratios (IRRs) for post-enrolment death during the first half of infancy between the ciKMC and the control arm using log-binomial generalized linear models with follow-up time in child months. We estimated the biologic interaction (i.e., interaction assessed on the additive scale) using the absolute excess rate due to interaction (AErI) for wealth status, infant’s sex, caste, religion, and mother’s literacy status [25, 26] using the appropriate interaction terms in the above-mentioned regression models. The regression analyses accounted for clustering of deaths among infants within the same household using robust standard errors.