6.10.1 Methodology
A conventional life table approach can be used effectively to discount the problems of censoring, but problems due to heaping remain in the data. The reported data on the duration of breastfeeding show that there is a clear pattern of heaping at various durations: first at three months and then later at multiples of six months. Such a pattern implies a classic example of digit preference or rounding off the durations to the nearest half year. It may also be the choice of breastfeeding duration conventionally accepted by the community. However, the pattem of heaping at digits, multiples of six months
5 In recent years dairy farms were set-up on most of the estates by the Government with external assistance to improve the income and nutritional level of the estate population. The majority of the estate population are Hindus, who in general tend to consume more milk than other community groups.
or the depressions at durations adjacent to these, observed among those who have stopped breastfeeding, is not present in the currently breastfeeding population (Figure 6.5), implying that reported durations are not accurate.
— Stopped breastfeeding — Currently breastfeeding
m 6 -
8 10 12 14 16 18 20 22 24 26 28 30 32 34
Months since birth
Figure 6.5. Distribution of reported breastfeeding durations and proportions of children being breastfed by months since birth.
Because of the possible contamination o f data due to such biases, alternative procedures are used only in the information relating to the responses to a single question, 'Are you currently breastfeeding?' to assess the breastfeeding pattem, so by passing the duration data which are deficient (Page, Lesthaeghe and Adegbola, 1977). The advantage o f such an approach is that it is largely free from the biases due to recall lapses and is less affected by any digit preference. These responses can then be linked to the age or the duration between the birth o f the child to whom the question refers and the date of the survey. If the age reporting o f children is satisfactory then this approach, which is known as the 'current status approach’, yields a reliable estimate on breastfeeding duration. Owing to the problems in the reported durations discussed earlier, instead of fitting the regressions using the traditional life table approach suggested by Cox (1972), this analysis makes use o f the current status data to arrive at
an estimate similar to the proportional hazard models.
Procedures for this approach have been suggested by Diamond, McDonald and Shah (1986). Briefly, they show that when the relationship between a binary outcome derived from current status data and a set of explanatory variables is examined, the analysis can be done by linking the complementary log-log transformation of the probability of occurrence of the event in the time interval parameter and those for the explanatory variables. By exponentiating the resultant maximum likelihood estimates, it is possible to estimate the relative risks which are equivalent to those derived from Cox's proportional hazard model.
In the application of this model to the SLDHS data the dependent variable is whether a child of a specific age is weaned at the time of the survey. The event of interest is assessed from the point of view of the child and not the mother; thus for the purpose of the model those who were weaned (failure event) were assigned 1, and others 0. One of the requirements of the application of this procedure is that the proportion of children breastfed should be negatively associated with the length of time since birth (or age of the child). Because of the sampling fluctuations for some ages (mainly arising from small numbers), this pattern could not be found in the data. Therefore, to meet the criteria, the ages of children (length of time since birth) were grouped in the following five categories: 0-3, 4-11, 12-17, 18-30, and 31-36. In their application of the model to Pakistan data, Diamond and others (1986) have also shown that grouping of age intervals does not unduly affect the estimates derived from the model. This grouping also conforms with the concept that the ratio of hazard functions for the two individual groups are constant at a given age or age group, which is a basic assumption of the proportional hazard models.
6.10.2 Application of proportional hazards model
For the model examining the probability of a child aged 0-36 months old currently being weaned the analysis was restricted to the most recent birth and all multiple births were eliminated. Normally, breastfeeding of a baby is abruptly stopped when a woman
discovers that she is pregnant causing an interruption to the event, although there is some evidence, albeit rare, of women continuing breastfeeding even during pregnancy
(Huffman et al., 1980). The SLDHS questionnaire was drafted such that current breast
feeding status was not asked about the elder children, if any, and it was assumed they were not being breastfed. Children of currently pregnant mothers were also eliminated from the model.6
The results of the proportional hazard model are presented in Table 6.7. The coefficients derived from the proportional hazards model presented here refer to the net- effects of each variable relating to the 'force of weaning' as opposed to duration of weaning, on the probability of a child being weaned at the specific age (Anderson, Rodriguez and Thome, 1984). Negative values of the coefficients in the model indicate that a child in that category is likely to be weaned later than the children in the baseline category. Similarly positive values indicate a high probability of weaning earlier than for children in the baseline category.
The individual relationships of the predictor variables with the outcome variable — whether a child at a specific duration since birth has been weaned or not — broadly correspond with the life table results discussed above, but the net effects of each of the variables are obviously different. Among the variables categorized under maternal fertility factors, only birth order is statistically significant. Sex of the child, which demonstrated a pattem of early weaning of boys, was found to exert a very weak association with durations of breastfeeding and is not a significant factor determining weaning. As shown at the bivariate level, the age of the mother at the time of the birth
6 Restricting the population to the last surviving birth in the birth cohort of three years before the