PropertyValue
is nif:broaderContext of
nif:broaderContext
is schema:hasPart of
schema:isPartOf
nif:isString
  • We used data from the District Level Household and Facility Survey 3 (DLHS-3), a nationwide household survey at district level, conducted in 2007–2008 in 34 Indian states and territories [19]. The DLHS-3 was designed as a cross-sectional study that used a stratified, systematic, multistage cluster sampling design [19]. The basic Indian vaccination schedule is proposed by Universal Immunization Programme (UIP). The UIP is the largest immunization program in the world and targets 27 million infants annually. The UIP protects children against 7 vaccine-preventable diseases: tuberculosis, diphtheria, tetanus, pertussis, polio, measles (added in 1985) and hepatitis (added in 1990). Vaccines are provided free of cost and delivered through strategies such as routine immunization, village health and nutrition days, and outreach campaigns [20]. In keeping with the definition in standard use in India, full immunisation is defined as a child 12–23 months of age receiving all of the following vaccines: a dose of BCG vaccine at birth (or as soon as possible); three doses of DPT vaccine at 6, 10 and 14 weeks of age; at least three doses of OPV at 6, 10 and 14 weeks of age; and one dose MCV at 9 months of age. Vaccination information of 12–23 month-old children in DLHS-3 was obtained either from health cards or from mother’s or caregiver verbal reports. We created two binary outcomes to study non-vaccination in this sample. First, children 12–23 months of age who had not received any of the following eight vaccine doses (1 dose of BCG vaccine, 3 doses each of DTP vaccine and OPV, and 1 dose of MCV were considered completely unvaccinated (CUV), and were compared to children who had received at least one dose of vaccine. Second, children were considered to have received no routine immunisation (no RI) if they had not received any of the five recommended doses administered only through routine services (1 dose of BCG vaccine, 3 doses of DTP vaccine, and 1 dose of MCV), and were compared to children who had received at least one routine immunisation dose. Full immunization coverage can be attained only through improving routine immunisation systems. For several decades, as part of the global eradication initiative, India has had a very strong polio programme operating largely in campaign mode in parallel to routine immunization services [21], [22]. We therefore also studied those children 12–23 months of age who had not received a single dose of vaccine from routine immunization services. Individual and household (compositional) characteristics: We included the following compositional variables: child sex (male or female), birth order (1, 2, 3, 4 and more), mother’s age (15–24, 25–34, or 35 years or older), mother’s and father’s educational attainment (0 year, 1–5 years, 6–8 years, 9–10 years, 11–12 years, or 13 or more years), caste (scheduled tribe, scheduled caste, other backward caste -OBC- and general), religion (Hindu, Muslim and others i.e. Sikh, Christian, Buddhist and others), antenatal care –ANC- (prenatal visits, tetanus injection during pregnancy), postnatal care (No PNC within 2 weeks), and household wealth. Household wealth index was computed by combining household assets and material possessions by IIPS and divided into quintiles (poorest to the richest groups accounting for the lowest to the highest quintiles). Contextual characteristics are defined at community, district and state levels. State-level characteristics considered included area of residence (urban and rural) and region of residence categorised into two groups as follows: The first group included Empowered Action Group States (EAG) and Assam (EAGA). The EAG states, which account for about 45% of India’s population and have particularly high fertility and mortality indicators, were designated as “High Focus States” by the Indian Government in 2001. Due to lagging social and demographic indicators, Assam is often considered with this group. EAGA states were: Assam, Bihar, Chhattisgarh, Jammu and Kashmir, Jharkhand, Madhya Pradesh, Orissa, Rajasthan, Uttar Pradesh, Uttarakhand).The second group (other states) included: Arunachal Pradesh, Manipur, Meghalaya, Mizoram, Sikkim, Tripura, Andaman and Nicobar Islands, Andhra Pradesh, Chandigarh, Dadra and Nagar Haveli, Daman and Diu, Delhi, Goa, Gujarat, Haryana, Himachal, Pradesh, Karnataka, Kerala, Lakshadweep, Maharashtra, Pondicherry, Punjab, Tamil Nadu, West Bengal).We used the term community to describe clustering within the same geographical living environment. Communities were based on sharing a common primary sample unit (PSU) within the DLHS-3 data as it is the most consistent measure of community in the DHS surveys [23]. Since poverty and education characteristics of communities were not directly available, they were constructed by aggregating individual-level characteristics at the PSU level. Specifically, these weighted measures were derived by summing the values obtained on individual women in each community and dividing then by the total number of women respondents living in each one. The community’s poverty status was defined as the proportion of households below 20% of wealth index. The proportion of women with no formal education was generated from native individuals in the database and aimed to represent female illiteracy in the community. In our study, these group-measures were based on an average of 3 women per community (from 1 to 31), which provides a sufficient number 1) to generate reliable estimates [24] and 2) to use Monte Carlo Markov Chains for achieving our computations [25]. The entire national sample (n = 65,617) of children aged 12–23 months was analyzed. Data typically have a hierarchical structure in which children were nested within mothers, mothers were clustered within households, households were nested within communities which were clustered within districts, and finally districts were nested within states. To account for unequal selection probabilities and ensure representativeness of the sample, we applied the appropriate sampling weights. Determinants of non-vaccination were assessed by using Bayesian binomial regression models. We specified a 4-level model for each binary outcome y, i.e., non-vaccination, for child i living in community j in district k and state l. Probability was related to a set of categorical predictors X and a random effect for each level by a logit-link function as logit (πijkl) = β0+βX+u0jkl+v0kl+f0l. A child level was defined by collapsing child-, mother- and household-level data. The linear predictor of the equation consisted of a fixed part (β0+βX) estimating the conditional coefficients for the covariates. The 3 random intercepts were respectively attributable to communities (u0jkl), districts (v0kl) and states (f0l), each assumed to have an independent and identical distribution and variance estimated at a corresponding level. All models were estimated by using Bayesian methods implemented via Markov Chain Monte Carlo (MCMC) simulation and the Metropolis-Hastings algorithm [26]. We used diffuse default prior distribution for all parameters [26]. Starting values of the distribution were derived from two previous estimations using Iterated Generalized Least Squares (IGLS) and second order penalised quasi-likelihood linearization (PQL2). MCMC estimation was adopted in the analysis to reduce bias in the estimates of random effect parameters. Indeed, such bias can arise when multilevel models with discrete outcomes are estimated using maximum-likelihood procedures [27]–[29]. All estimations were performed by MLwiN within STATA 12 MP (Stata, Corp.) and MLWiN 2.26 through runmlwin procedure [30]. Our computations were based on chains of length 50 000 iterations after a burn-in of 5000. Bayesian deviance information criterion (BIC) was used to estimate the goodness of fit of consecutive models [31] The BIC values for each model were compared, and the model with the lowest value was considered the better one for hierarchical models [31]. We examined separately the association between non-vaccination and compositional (individual-household) and contextual variables. The first model is a null model (Model 1), which provides information on the extent to which communities, districts and states vary and further justify assessing random effects at these levels. Model 2 included only individual characteristics while model 3 contained community characteristics. Model 4 expanded model 3 by adding individual level variables. We further fitted a fifth model to analyse a cross-level interaction between household wealth and area of residence (rural and urban). Since we found a significant (p = 0.035) interaction term (area of residence * wealth index), we present separate models including all individual and contextual variables stratified for rural and urban areas of residence. The fixed effects, i.e., the association between non-vaccination and selected variables, were shown as odds ratio (OR) with its 95% credible interval (CrI). Meanwhile, random effects (measures of variation) were estimated by median odds ratio (MOR) rather than using intra-cluster correlation (ICC) which is better fitted for linear models [32], [33]. The MOR quantifies the unexplained contextual heterogeneity, otherwise it quantifies contextual-level variance on the odds ratio scale and is always greater than or equal to 1 [32]. This study is based on an analysis of existing survey data with all identifier information removed. The survey was approved by the Ministry of Health & Family Welfare, Government of India and the International Institute of Population Sciences (IIPS) institutional review board. All study participants gave informed consent before participation and all information was collected confidentially. Data of DLHS-3 were obtained from the IIPS as they are made available in the public domain for analysis by researchers. Therefore, no additional ethics review is required for this work by the Montréal University committee of ethics.
rdf:type