This study used the 2003 National Demographic and Health Survey, a cross sectional national survey implemented by National Statistics Office. Face to face questionnaire-guided interview was used to collect health and demographic related information from 13, 633 women of reproductive age. Three-stage sampling design was used to represent the country in seventeen regions. In the first stage, 819 villages in each region were selected using the 2000 Census master list. In the second stage, enumeration areas were selected in each village. An enumeration area is defined as an area with discernable boundaries and 150 neighboring households. In the third stage, an average of 17 households was selected from each enumeration area using systematic sampling. Actual data collection was carried out from June to September 2003 with a 99% response rate [5].
Secondary data that was formally requested from Macro International was used to answer the objectives of the study. Since no patient intervention was involved, there was no need for the study to undergo ethical approval. The respondents are identified through unique codes and neither by their names nor addresses to ensure confidentiality of respondent's identity.
Analysis was restricted to last-born children aged 6–36 month old at the time of survey. Children less than 6 months old (n = 639) were excluded since their breastfeeding duration cannot be determined, while children more than 36 months old (n = 1,237) were also omitted in the analysis to minimize possible recall bias. Dead children were included as long as their age of death satisfies the age criterion. Thus, out of 4,920 last-born children surveyed, only 3,044 children were eligible for inclusion.
Pregnancy intention was categorized into three categories: (1) wanted: mother wanted to become pregnant then, (2) mistimed pregnancy: mother want to wait later and (3) unwanted: mother did not want any more children. Mothers who responded “do not know” were excluded from the study. There were two breastfeeding outcomes in this study: breastfeeding duration and initiation. Mothers’ of children reported less than six months of breastfeeding were considered as having short breastfeeding duration, otherwise they were categorized as having prolonged duration of breastfeeding. On the other hand, mothers who answered less than one hour were considered as having initiated breastfeeding early otherwise were considered late.
Maternal socio-demographic, prenatal and delivery-related factors were considered as potential confounders. Maternal age at delivery was determined. Education was categorized as no education/primary, secondary and higher. Marital status was categorized as single (previously married) of currently married (including living-in together). Maternal employment was grouped into not working, working all year and working on a seasonal basis. Type of residence was categorized as either urban or rural. Type of delivery was categorized as normal vaginal or cesarean section. In this study, parity and socio-economic status were also assessed as possible effect measure modifiers. Parity was categorized as primaparity (only one live birth) and multiparity (2 or more live births). Socio-economic status was determined by principal component analysis. Ownership of seven household assets, namely, refrigerator, motorcycle, television, radio, electricity, automobile and radio were subjected to principal component analysis to derive socio-economic index. Mothers with zero or less socio-economic scores were considered poor otherwise categorized as non-poor [10].
Figure 1 shows the different stages to determine the independent association between pregnancy intention and optimal breast feeding practices, namely, (1) bivariate analysis (2) stratified analysis and (3) multivariate analysis.
Bivariate analysis was used as a preliminary step in screening for the probable confounders. To be more conservative in the of possible confounders, those with a p-value less than 0.25 in at least one category were considered as probable confounders [11]. In the association between pregnancy intention and breast feeding initiation, place of residence, maternal education, employment, marital status, maternal age of conception and type of delivery variables were assessed as probable confounders. In the association between pregnancy intention and breastfeeding duration, same variables were considered as probable confounders in addition to breastfeeding initiation. Stratified analyses were also performed to determine if socio-economic status and parity were probable effect measure modifiers (EMM). A chi-square test of homogeneity with a p- value less than 0.15 qualifies a variable as a probable effect measure modifier [11].
Multivariate logistic regression analyses were performed to determine if EMM screened in stratified analysis and the probable confounders screened in the crude analysis were true effect measure modifiers and confounders. The significance of EMM screened in the stratified analysis were determined using Likelihood Ratio Test. Interaction terms were retained if the p-value generated was less than 0.10, indicating that the examined variable is indeed an effect measure modifier.
After assessment of effect measure modification, probable confounders identified during crude analysis were also assessed. The process of assessing confounders started with building the full model. The full contains the outcome variable (breastfeeding duration/initiation), the exposure variable (pregnancy intentions), all probable confounders and interaction terms. The probable confounder with the largest p-value was the first variable to be removed from the full model. In the presence of effect measure modifier as determined by the LR Test, stratum specific assessment of confounding was performed. Confounding was determined by comparing the odds ratio in the model with the deleted variables to that of the full model. If the odds ratio of the full model which contains the exposure, outcome and probable confounders differs from the OR of the reduced model by 10% or more in any of the stratum of the effect measure modifier, then the variable was considered a confounder.