Mothers' education but not fathers' education, household assets or land ownership is the best predictor of child health inequalities in rural Uganda

Study area

The study was carried out in Hoima district, which is located in western Uganda and shares a border with the Democratic Republic of Congo along Lake Albert. The district has a population density of 80 per square kilometre with a total population of approximately 370,000, comprised of mainly peasant farmers of the Banyoro tribe [21]. The district has fertile soils and enjoys a bimodal rainfall pattern ranging from 800 – 1500 mm per annum enabling two harvests a year. Farming activities include cattle rearing, growing tobacco and coffee for cash; and maize, cassava, potatoes, beans and groundnuts for consumption.

Design and sampling

The present survey utilised a two stage probability proportional to size cluster design [22]. Data was collected with respect to infant- and young child feeding knowledge and practices, anthropometry and indicators for household socio-economic status. The sample was made large enough to estimate a prevalence rate of wasting similar to that assessed at the national level (4.0%), at 95% level of confidence [23], with 2% precision error and an assumed intra-cluster correlation coefficient of 0.11. The national census data [21] was used to estimate the population under-2 years of age in Hoima district (24,000). A sample size of 720 children was calculated based on the formula by Cochran [24], using the SampleXS software. At the first stage of sampling, a total of 72 villages were selected on probability proportional to population size basis. At the second stage, a total of 10 households were systematically sampled from each village selected at the first stage, providing a self-weighting sample with each child/mother pair in the district having an equal probability to be selected into the sample. A household was defined as a group of people living, cooking and eating together. One child under 2 years was enrolled per household after an informed verbal consent of the mother/caregiver. In case a household had two children in the target age group or twins one was selected randomly. Children with major handicap, disability or malformation were excluded from the sample. Problems encountered with sample selection included the existence of newly formed villages hitherto unrecorded in the census register. In one case an old village had been split; one of the "offspring" villages was selected randomly.

Data Collection

A structured questionnaire constructed in English and translated into Runyoro, the local language, was pilot tested and updated in survey similar settings and administered face-to-face to mothers/caregivers in their home settings. Trained non-medical university students gathered data in September and October 2002. To minimise bias, training of interviewers emphasized proper and random identification of respondents, eligible children, and questionnaire administration. A village leader followed data collectors through the village, and traditional village protocol was observed.

Information about the educational level of parents, household durable assets, materials of the dwelling structure and land ownership was collected as proxy makers of household socio-economic status. Child's age was obtained through birth certificates, health cards, or recall using a calendar of local events while taking precautions to minimize field errors [25]. Recumbent length measurements were taken with specially designed length boards that measured to the nearest 1 mm following standard recommendations [26]. The questionnaire was based on the demographic and health survey (DHS) questions [27] to avoid separate validity and reliability checks of the data collection instrument.

Measurements

The dependent variable (stunting) was constructed using child length, age and sex, and was defined as length less than minus 2 standard deviations (<-2 SD) from the median of the NCHS/WHO reference population [5].

The education level of fathers and mothers was assessed on scales ranging from (1) never went to school, to (7) college/university. Two categorical variables were constructed yielding four categories for mothers: (1) no formal education, (2) stopped in primary, (3) completed primary (minimum 7 years of school) and (4) stopped above primary. Five categories were yielded for fathers because they were more educated: (1) no formal education, (2) stopped in primary, (3) completed primary, (4) stopped in secondary, and (5) completed secondary 4 or above (minimum 11 years of school).

Household durable assets (cupboard, hurricane lamp, radio, bicycle, fuel, boat, telephone, refrigerator, motorcycle and car) were assessed as (1) available/ in working condition and (2) not available/ not in working condition. Four components of the dwelling structure were assessed including number of rooms, roof – (1) thatched, (2) corrugated iron sheets or tiles; floor – (1) mud, (2) cement; wall – (1) thatched, (2) mud and pole, (3) unburned bricks, (4) burnt bricks built with mud, (5) burnt bricks built with cement and (6) cement/concrete blocks. A household wealth (asset) index was constructed from 15 variables (household durable assets and dwelling structure) using principal components analysis [28]. The index scores were divided into quintiles (1) poorest, to (5) least poor.

Land is usually given special socio-economic status in many subsistence agricultural communities in Uganda [29]. In order to assess for its independent contribution to child well being land ownership was analysed independent of other household assets. It was assessed as a continuous variable using an open question "How much land does your family have for agriculture or any other activity?" A football pitch was used to estimate the size equivalent to one acre. A categorical variable was constructed in terms of; (1) no land at all, (2) one acre, (3) two acres, (4) three acres and (5) four or more acres.

Child age in months was assed as a continuous variable and a categorical variable was constructed yielding four categories: (1) 0–5 months, (2) 6–11 months, (3) 12–17 months, and (4) 18–23 months. Child sex was ordinal (1) male, (2) female.

Analysis

Data was entered in Epidata software [30], and was analysed in epi info (version 6.04d) and SPSS (version 12.0). Of 720 children recruited, 23 (3%) had missing data on anthropometric measurements or were extreme outliers and were dropped from the analysis of stunting. First, socio-economic indicators (mothers' and fathers' education, household wealth index and land ownership) were evaluated for the extent of shared variation by obtaining bivariate correlation coefficients (Spearman's rho). Second, unadjusted associations of stunting with socio-economic and demographic variables were examined by use of odds ratios. Third, backward conditional logistic regression was done simultaneously controlling for socio-economic and demographic factors. The pattern of stunting with sex across different socio-economic strata was evaluated with chi-square tests and bivariate logistic regressions. Mean age differences between boys and girls were assessed by the student t-test. Tests for trend across categories were performed by treating the categories as continuous variables in the logistic regression analyses. The clustering effect was not controlled for as the large number of primary sampling units (72) in the study minimizes it. The goodness-of-fit for regression models was assessed with Hosmer and Lemeshow test; with the null hypothesis that the model adequately fits the data (significant chi-square p-value implies that the model does not fit).

Ethics

Approval of the study was granted by Makerere University Faculty of Medicine Ethics and Research committee, the Uganda National Council for Science and Technology and the Regional Committee for Medical Research Ethics, West Norway (REK vest).

Stunting distribution by socio-economic status

Unadjusted logistic regression analysis indicated that 3 socio-economic indicators (fathers' education, mothers' education and household wealth index) had a graded pattern with stunting across strata with the least educated or most poor being more likely to have stunted children. The corresponding p-values for test of trend was <0.001, 0.03 and 0.01 for mothers' education, fathers' education and household wealth index, respectively (Table 1). Land ownership however, showed no differentials with stunting across strata.

The extent of association between socio-economic indicators was examined. The correlations (Spearman's rho) between mothers' and fathers' education was 0.50; mothers' education and household wealth index = 0.37; mothers' education and land ownership = 0.12; fathers' education and household wealth index = 0.39; fathers' education and land ownership = 0.14; and household wealth index and land ownership = 0.18. Their colinearity was sufficiently low to allow them into the same model.

The four socio-economic indicators (mothers' education, fathers' education, household wealth index and land ownership), child age and sex were simultaneously entered in a backward conditional logistic regression. Only mothers' education, child age and sex appeared in the last step model. Children belonging to mother with no formal education or to mothers who stopped in primary school were significantly more likely to be stunted compared to their counterparts with mothers who were educated beyond primary school (OR = 2.1, Table 1). Being a boy (OR = 2.0) or older in age was also significantly associated with stunting.

Stunting distribution by sex and socio-economic status