Research design and study population
The monitoring and impact evaluation of the JP utilized a ‘before-and-after’ non-experimental type of study design. A baseline survey was carried out in early 2011 to establish the baseline levels of child undernutrition and prevalence of optimal breastfeeding practices. After this, the interventions were implemented from April 2011 to September 2012, shortly after the Baseline Survey was conducted. The Endline Survey was conducted in late 2012 to assess the effectiveness of JP interventions in improving the prevalence of optimal breastfeeding practices and in reducing undernutrition among the target children. Both baseline and endline surveys for outcome and impact evaluation were conducted by one of the authors (OPS). To examine the association between the JP interventions, specifically exposure to peer counselors and mother’s membership in breastfeeding support groups, and EIBF and EBF, data from the Endline Survey were used in this paper.
There were six JP sites throughout the Philippines representing one highly urbanized city and one rural municipality for each major island group: the cities of Naga, Iloilo, and Zamboanga, and the municipalities of Ragay in Camarines Sur; Carles in Iloilo; and Aurora in Zamboanga del Sur. For both surveys, a stratified two-stage systematic random sampling was employed to select the study participants. Each JP site served as a stratum. In the first stage of sampling, barangays (i.e., Philippine term for a village) or contiguous small barangays with a minimum of 600 households were randomly selected and served as primary sampling units (PSUs). The 2010 Philippine Census was used as the sampling frame for the PSUs. The PSUs in each city/municipality were selected systematically with probability proportional to the PSU’s population size. Prior to sampling for the Endline Survey, 44 barangays in Zamboanga City were excluded because of the deteriorating peace and order situation in the area. From each remaining PSU, children less than two years old were randomly selected, with equal probability, using the lists of eligible children in each barangay. The lists of eligible children from each barangay were collected from local health workers and validated and updated for completeness by a team of mappers/validators.
A pre-tested, structured, paper-based interview schedules were developed by the investigators. These interview schedules, which were originally designed in English, were translated to the local languages (e.g. Tagalog, Bicol, Hiligaynon, Bisaya, Chavacano, and Tausug) spoken in the different study sites. The data collectors were trained in administering the interview schedule. After obtaining informed consent, face-to-face interviews using the localized version of the interview schedules were conducted to collect socio-economic data from the sampled households, demographic characteristics of the mothers, and their IYCF practices for the index child.
For this survey, a sample size of 2,584 mother-infant dyads from all six JP sites was required to detect a 3% absolute decrease, from the baseline, in the prevalence of underweight for age, using a level of significance of 0.05, 80% power, a design effect of 1.2, and an allowance of 10% non-participation. The Endline Survey was similar with the Baseline Survey, but additional questions on whether the respondents had received JP interventions were included. For this paper, only mothers who were the actual caregivers of the target children were included in this analysis. Data from other child caregivers, such as fathers, were excluded in the analysis to minimize the effect of different quality of recall that may arise from having different types of child caregivers.
Operational definition of study variables and description of interventions of interest
The outcome variables in this study are EIBF and EBF. For EIBF, mothers were asked in the original interview schedule how long after birth was the index child first put to the mother’s breast, and the responses were recorded in number of hours. For this analysis, we recoded this variable into late initiation (after one hour of birth) as the baseline, or early initiation (within one hour of birth) of breastfeeding. For EBF, mothers were asked at what age, in months, were solid foods or other liquids were first introduced to the infant. We dichotomized this variable into mothers who gave solid/liquid foods to the child six months after birth and mothers who did not.
The exposure variables in this study were home visit/s by a peer counselor during the prenatal period, home visit/s by a peer counselor after delivery, and membership of the mother in a breastfeeding support group during the index pregnancy. Being visited by a peer counselor during prenatal period and membership of the mother in BF support group during the index pregnancy were considered as exposure variables in the model for EIBF, while these two variables, together with home visit/s by a peer counselor after delivery were considered as exposure variables in the model for EBF. All three are dichotomous variables (either visited by a peer counselor or not, or member of a support group or not.) Mobilization of peer counselors and breastfeeding support groups were among the interventions implemented by the JP. Peer counselors were volunteers, often female local community workers, who were engaged and trained to educate mothers about EBF, correct positioning and attachment of the baby during breastfeeding, visit pregnant and post-partum mothers at home to advocate breastfeeding, orient pregnant mothers about the Milk Code which prohibits advertisements of breastmilk substitutes, and teach pregnant mothers how to prepare for their delivery. These peer counselors were trained for five days by the Program staff (three days to administer the interview schedule and two days of supervised mock interviews in the field) and asked to cover a number of households in a barangay (i.e., village) or part of a large barangay. They were expected to provide contact details of pregnant mothers or mothers who have recently delivered. Meanwhile, breastfeeding support groups were led by female influential community members who engaged target mothers in small group discussions about many topics related to breastfeeding including the benefits of breastmilk, advantages of breastfeeding, correct breastfeeding techniques, proper diet of lactating mothers, and complementary feeding. Like the peer counselors, the leaders of breastfeeding support groups were likewise trained by local Program implementors.
The probable confounders in these associations of interest included place of residence, age of mother in years, total monthly household income, employment status of mother and partner, number of people living in household, number of living older siblings, mode of delivery of index child, birth attendant of index child, place of delivery, gender of child, maternal knowledge score, attendant during prenatal services, month when prenatal service was first availed, and membership in the Pantawid Pamilyang Pilipino Program (4Ps). The 4Ps program is a conditional cash transfer program implemented by the Philippine government which targets economically-disadvantaged Filipino families in return for complying with set conditions on children’s education and the family’s utilization of health services such as prenatal check-up and child vaccination (35). In the model for EBF, EIBF was also considered as a probable confounder.
Data management and analysis
Data quality checks, such as checks for duplicates and range checks were performed on the dataset prior to any analyses. Some quantitative variables, such as age and monthly income, were recoded to allow the assessment of possible linear trends in the association between these variables and the outcome (36). Some categorical variables were recoded to ensure that estimates for subsequent regression analyses would be stable. Other categorical exposures, such as place of residence, marital status, birth attendant, and mode of delivery, were recoded to ensure that each stratum would have sufficient number of observations. Maternal knowledge score was aggregated from seven yes-no questions. Incorrect answers or “don’t know” answers were coded as incorrect and given a ‘0’ score, and a score of ‘1’ was given for each question answered correctly. Scores may range from 0 to 7, with higher scores implying better maternal knowledge. Lastly, variables that were thought to be correlated were combined, such as marital status of the mother and employment status of partner.
After the data management procedures described above, the dataset was declared as survey data. The sampling weights and strata (i.e., the six study sites) were also defined. All subsequent analyses, except for non-parametric tests, were weighted. However, the counts presented in the Results section were unweighted. No observations were deleted at any point in the analysis to ensure that standard errors can be computed correctly.
The distributions of continuous variables were presented using appropriate measures of central tendency. Frequencies and proportions were used to describe the distribution of categorical data. For descriptive statistics, weighted proportions were estimated; however, the counts were not weighted. The aforementioned exposure variables were cross-tabulated with each of the two outcome variables and the association of each of these exposures with each outcome variable were tested with Pearson’s χ2 test for categorical exposures, the adjusted Wald test for normally-distributed continuous variables, or the Wilcoxon rank-sum test for skewed continuous variables. The distribution of missing data was shown for each variable, but they were not included in estimating the p-values. For each of these associations, crude odds ratios (cOR) were estimated using simple logistic regression for survey data. The cORs and the p-values for these cross-tabulations were noted.
As part of screening potential confounders, each probable confounder and the outcome variable were cross tabulated with each of the exposure variables using similar statistical tests as described above. Likewise, the cORs and the p-values for each of the cross-tabulations were noted. Probable confounders with strong evidence of association with the outcome, as well as with any of the exposures, but were not in the causal pathway of the other variables, were fitted into the final logistic model. At this point, observations with missing data for any of the variables of interest were excluded from the analysis.
In this paper, we ran two logistic regression models, one for each outcome variable. In building the final models, the main exposure variables were fitted first. Afterwards, variables meeting the operational definition of a confounder as described above, were fitted into the model, starting with the variables with the smallest p-value in the cross-tabulations with the outcome, and so on. After this, any variable deemed to be an important confounder based on the literature (even if they have not shown any strong association with the exposure or outcome in this dataset) was forced into the model starting with the variables with the smallest p-value in their respective cross-tabulations with the outcome, and so on. Any remaining variables were fitted into the model one by one, starting with the smallest p-value in their respective cross-tabulations with the outcome. If any of these variables changed the estimate of the OR for any of the main exposure variables by >10%, then they were retained in the final model; otherwise, they were excluded.
Once grouped quantitative variables were fitted into the model, test for departure from the linearity assumption was carried out by observing the stratum specific odds ratios (OR), and by doing an adjusted Wald test. If the test for departure from the linearity assumption was statistically significant, or the stratum-specific ORs did not show evidence of a linear trend, stratum-specific ORs were presented. Otherwise, a common estimate for the linear effect of the exposure variable on the outcome was reported (36). After testing for departure from the linearity assumption of grouped quantitative variables, the interaction between the exposure variables were assessed (37). Any significant interaction parameters were shown, and linear combinations were used to estimate interaction parameters.
A level of significance of 0.05 was used in all analyses, and 95% Confidence Intervals (CIs) were reported. Data management and analyses were carried out in Stata 14.2 (38).