Quantitative analysis

A baseline census was conducted in all clusters in 2004, prior to the start of the intervention. All women aged between 10 and 49 years were enumerated and a random sample of 104 women aged between 17 and 43 years per cluster was then drawn to be interviewed for two follow-up quantitative surveys as part of this secondary study. Sampled women (‘main respondent’ hereon) were visited to complete the first follow-up in November 2008 to March 2009, and a second follow-up in October 2009 to January 2010.

Each follow-up survey contained questions about the size of the extended family of the main respondent and her husband (those alive and those in the village), the health of all household members, food and liquid intake of children aged under 6 years, knowledge about child nutrition, intervention participation (in treatment clusters) and socioeconomic variables such as adult work. The height of the main respondent and the height and weight of children under 6 years were also collected by trained enumerators.

The main outcome for our analysis is the child HAZ score, which is a long-term indicator of health that reflects nutrition and morbidity since birth, and should be sensitive to any effects of intervention exposure in early life. It is calculated by comparing the height of the child with the median height in the WHO reference population of children of the same gender and age in months.³⁵

The sample was balanced between treatment and control clusters along a range of variables collected at baseline¹⁷ (table 1). The baseline characteristics of the two groups remained similar even after accounting for attrition between the baseline and first endline survey, indicating that randomisation was not jeopardised (table 1).

Distribution of household and women characteristics in controls and differences with treatment group

Household and mother level characteristics in 2004 corresponding to married main respondent mothers present in the second follow-up survey with children born after the intervention began in July 2005.

*p<0.1, **p<0.05, ***p<0.01. P values are calculated using the wild cluster bootstrap t procedure described by Cameron et al.³⁷

†Continuous variable, for which the mean is reported.

‡ Binary variable, for which proportions are reported.

For this analysis, we use a sample of children who were born since July 2005; and whose mothers are married main respondents in the follow-up surveys (80% of the sample). This sample selection ensures that we measure effects on children whose mothers were eligible to receive visits from a peer counsellor; and allows us to compare effects of the mothers’ relatives with those of her husband. Children in the estimation sample were aged between 0 and 53 months at the time of the endline surveys. Online supplementary appendix 1 provides a timeline of the original trial and the quantitative data collection, and of our sample inclusion criteria.

bmjopen-2017-019380supp001.pdf

Table 1 presents the means of basic demographic and socioeconomic characteristics for women in the analysis sample living in control clusters at baseline, the differences in the means between the control and treatment groups and the p value of this difference. The last two columns allow us to assess whether the randomisation holds in our selected sample. Women assigned to the control group were 24.6 years on average; 71.8% were married and while 70.1% had completed at least primary education, only 7.6% had completed secondary education. In line with the general profile of communities in Mchinji, 95.4% of sampled women were Chewa ethnicity and 98.3% were Christian. The average household size was 5.6 members and all households were engaged in agricultural activity.

Table 2 displays statistics on the size of extended family networks of the children in our analysis sample. Most children have their grandmothers alive (87.3% have maternal grandmothers alive and 80.7% have paternal grandmothers) and their parents have a relatively large number of siblings, with an average of more than two brothers and two sisters each.

Distribution of family networks indicators in controls and differences with treatment group for sampled children.

Sample includes all children born since July 2005, who were aged 0–53 months at the time of interview, and whose mothers were married main respondents to the follow-up surveys in 2008–2009 and 2009–2010. A pooled dataset from both follow-up surveys is used to construct means.

*Binary variable for which percentages are reported.

The quantitative analysis aims to determine how different family members influence the effectiveness of the infant feeding intervention. Before estimating the main model, we study the relationship between baseline characteristics of mothers and their households and measures of the extended family using linear regression. Table 3 reports these results. It indicates that children whose grandmothers are alive have on average younger mothers, who are more likely to have completed at least primary education, less likely to be working as farmers in 2004 and are from more socioeconomically advantaged households, as measured by a composite wealth index constructed using principal components analysis as recommended by Filmer and Pritchett.³⁶

Relationship between baseline characteristics and family network size, with p values

Ordinary Least Squares regressions with baseline characteristics gathered in 2004 as the dependent variable and family networks as independent variables. Sample contains married main respondent mothers present in the second follow-up survey with children born after the intervention began in July 2005. SEs computed using the cluster-correlated Huber-White estimator are reported in parentheses and p values are also reported in parentheses. P values are calculated using the wild cluster bootstrap t procedure described by Cameron et al. ³⁷ The wealth index was calculated using principal components analysis as recommended by Filmer and Pritchett.³⁶

*p<0.10, **p<0.05, ***p<0.01.

Our main specification is the following linear multivariate regression:

where HAZij is the height-for-age z-score of child i in cluster j. T_j is a treatment exposure indicator, which captures whether the child was born to a mother living in 2004 (pre-intervention) in a cluster that was assigned to receive the programme. We therefore use an intent-to-treat estimator. Maternal_grandmother_ij and Paternal_grandmother_ij are binary variables indicating, respectively, whether the maternal and paternal grandmother is alive. Total_mothers_siblingsij (Total_fathers_siblingsij) captures the total number of siblings of the child’s mother (father) who are alive. We use two definitions of this variable in different specifications of the model: (i) brothers and sisters (separately) of each parent and (ii) the total siblings of each parent. X_ij and Z_j are vectors of control variables at the individual and cluster level, respectively. These include all baseline characteristics where significant differences between households with different extended family members alive were detected, and interview month and year indicators to account for month-year-specific shocks. We do not adjust the data for missing information.

We fitted three models, one crude model with the intervention term only, and two full models as specified in the equation each treating parent siblings differently.

The coefficient β captures the effect of the programme for children whose maternal and paternal grandmothers are dead, and whose parents are only children, while the coefficients β2, β4, β6 and β8, represent the effects of the extended family members on HAZ scores in the control group. The coefficients β3, β5, β7 and β9, associated with interaction terms between variables capturing extended family relations and the indicator for programme allocation, estimate the additional effect of the programme for children with different types and numbers of extended family members. A positive (or negative) significant interaction provides evidence that the programme effect is enhanced (or diminished) in the presence of that particular family member.

Errors ε_ij are assumed to be uncorrelated between individuals in different clusters but are allowed an unrestricted correlation structure within clusters. To account for correlation within clusters, SEs must be adjusted to prevent downward bias, and incorrect inference. Given the small number of clusters in the study (12 intervention and 12 control clusters), we adopt wild cluster bootstrap methods as recommended by Cameron et al.³⁷ Associated 95% CIs can be calculated using a computationally intensive method suggested by Colin Cameron and Miller.³⁸ The bootstrap adjustment applied here was studied in detail by Fitzsimons et al and was found to perform well.¹⁷ Data from both follow-up surveys are pooled to improve statistical power.

The extended family network is defined according to which members of the family are alive, rather than which ones live in the same village or household. This is in case treatment exposure affected decisions over where to live, which would cause a measure of family network size based on residence to be correlated with the intervention and thereby bias estimates. The benefit of defining the size of the family network according to which members are alive is that this is almost certain to be invariant to programme exposure.

We choose to define T_j by exposure to the intervention rather than actual participation since participation in the programme was voluntary and also relied on the ability of peer counsellors to locate eligible women. Women who peer counsellors did not manage to trace or who chose not to take part in the programme may be different from those who did participate. The existence of such systematic differences would potentially introduce some unobserved correlation between the treatment interaction variables and HAZ scores if T_j were defined on the basis of actual participation. Indeed, Fitzsimons et al report that women who received the visits tend to be poorer.¹⁷ Defining treatment based on residence at baseline rather than at the time of the follow-up interviews also alleviates concerns of bias in case there was purposeful migration into treated areas by control-group assigned households.