Country characteristics and acute diarrhea in children from developing nations: a multilevel study

We designed a cross-sectional, transnational and multilevel study that used level-1 data (child, mother and household characteristics) from the Demographic and Health Survey (DHS) phase-V [14] and level-2 data (country characteristics) from the World Bank (WB) country data [15].

The DHS phase-V collected data from 41 developing countries from 2004 to 2010. A nationally representative, probabilistic sample including rural and urban areas was collected from each participating country. Respondents were selected through a multistage, stratified sampling procedure of households. Between 5,000 and 30,000 households were surveyed per country. Data was gathered for the following countries: Albania 2008–2009, Azerbaijan 2006, Bangladesh 2007, Benin 2006, Bolivia 2008, Cambodia 2010, Colombia 2010, Congo 2005, Egypt 2005–2006, Philippines 2008, Ghana 2008, Guyana 2009, Haiti 2005–2006, Honduras 2005–2006, India 2005–2006, Indonesia 2007, Jordan 2007, Kenya 2008–2009, Lesotho 2009, Liberia 2007, Madagascar 2008–2009, Malawi 2010, Maldives 2009, Mali 2006, Namibia 2006–2007, Nepal 2006, Niger 2006, Nigeria 2008, Pakistan 2006–2007, Peru 2004–2008, Democratic Republic of Congo 2007, Dominican Republic 2007, São Tomé é Príncipe 2008–2009, Sierra Leone 2008, Swaziland 2006–2007, Tanzania 2010, East Timor 2009–2010, Ukraine 2007, Uganda 2006, Zambia 2007 and Zimbabwe 2005–2006. We excluded Ukraine from the analysis because the country did not apply the child health module of the survey. Information from the remaining 40 countries was merged to create a single dataset, which included 395,485 households with children. The dataset was further limited to biological mothers answering the survey to assure comparability (384,662), living children (359,527), permanent household residents (349,849) and cases with complete information (348,706).

After careful analysis, we concluded that the WB country data was the best source of level-2 data in this study because of its country comparability and robustness when compared to data from other sources. These data included the 2010 indicators: per capita gross domestic product (GDP), Gini-coefficient, and health expenditure as a percentage of GDP.

We divided the variables according to the data source: level-1 variables included the child, mother and household characteristics, and level-2 variables included country data. Initially we considered three levels of analysis –child, household and country- but most households had only one child under the age of five, so it was decided to include only one child per household, the youngest, and conduct a two-level analysis.

Level-1 data, children: sex coded as 0 = male and 1 = female, age of child in months, immunization defined as completeness of WHO schedule, coded as 0 = incomplete and 1 = complete, duration of breast feeding in months, possession of health card coded as 0 = no and 1 = yes, undernutrition defined as a BMI (body mass index) under the 5th percentile, and birthweight codified in dummy variables as follows: normal -above 2,500 g- 0 = no, 1 = yes; low -below 2500 g- 0 = no, 1 = yes; and no weighted, 0 = no, 1 = yes. Although it is not ideal to include a birthweight missing indicator, taking into consideration that 46 % of the children did not have this information, imputation was ruled out.

Level −1 data, mother: age in years, education: no education coded as 0 = no, 1 = yes; elementary education coded as 0 = no, 1 = yes; high school education coded as 0 = no, 1 = yes; superior [technical or professional] education coded as 0 = no, 1 = yes), current employment coded as 0 = no and 1 = yes and planned pregnancy referring to the pregnancy in which the index children was the outcome and coded as 0 = no; 1 = yes.

Level −1 data, household: number of household members defined as number of people living in the same home, place of residence coded as 0 = rural and 1 = urban, nuclear family defined as a social unit conformed exclusively by two parents and one or more children: 0 = non-nuclear family, 1 = nuclear family, sanitation score based upon water source and waste disposal, both classified as improved or unimproved, from zero to two, with higher grades indicating better sanitation [16], and wealth index calculated by the DHS considering income, possessions and quality of life, with higher grades indicating greater wealth. Wealth index ranges from 1 to 5.

Level-2 data, country: Country wealth coded as a set of dummy variables as follows: Low income (1 = gross domestic product per capita (GDPpc) of US$1,025 or less), Lower middle income (1 = GDPpc between US$1,026 and US$4,035), Upper middle income (1 = GDPpc between $4,036 and $12,475), and High income (1 = GDPpc of $12,476 or more) [15], Inequality based on the Gini coefficient (1 = top 25 % unequal countries and 0 = more equal countries) and health expenditure coded as a set of dummy variables based on the percentage of GDP expended on health. Low health expenditure (1 = 5 % or less), Middle health expenditure (1 = between 5.1 and 10 %), High health expenditure (1 = more than 10 %). We considered in the initial models country homicide rates and total country population, but these variables were omitted in the final models because of their lack of association with diarrheal disease and their negative effects on the model’s validity, measured using residual files and reliability estimates.

The analysis was conducted considering known factors associated with diarrhea and the country characteristics to study. Multilevel analyses were preferred because the hierarchical nature of the data violated the principles of independence and homogeneity required for a single-level analysis [17]. Country characteristics were shared by many children.

All the variables were used in the initial analysis. Afterwards, sanitation and number of members in the household were excluded due to co-linearity. The wealth index was built considering both variables. Possession of a health card and undernutrition were excluded for their co-linearity with immunization and wealth index, respectively. Although low BMI has been highly associated with diarrhea, wealth index was used over undernutrition in the models because the former had 35 % of its data missing.

The statistical analysis was performed using SPSS 20.0 (IBM) and HLM 7 (Scientific Software International, Inc.), as follows: 1) we merged the individual datasets of the 40 countries, 2) we calculated descriptive statistics for categorical (proportions) and numerical variables (mean, standard deviation, minimum and maximum values), 3) we obtained bivariate odd ratios using hierarchical linear modeling logistic regressions of diarrheal disease in all of the studied variables, and 4) we generated multivariable models for diarrhea using hierarchical linear modeling. Stepwise multilevel logistic regression equations were estimated. Individual, family and household factors were included as possible predictors of diarrhea, and differences were deemed to be significant with P-value less than 0.05. The large sample size allowed us to find small differences with narrow 95 % confidence intervals. Finally, multilevel modeling was used to explore the association of country characteristics with diarrheal disease (between countries associations) adjusting for individual, family and household predictors of the condition (within countries associations) [18, 19]. Full maximum likelihood was used to fit the models. Random effects were estimated only for indicators with variations between groups that could be explained by the studied variables, allowing the coefficients to vary across groups. Those level-1 indicators were centered on the country mean to avoid the problem of co-linearity. All other variables, as well as the neighborhood variables, were centered on the grand mean and we constrained their variance. The final model can be seen in Appendix 2. We have calculated median odd ratios (MORs) and intra-class correlations (ICC) for the models, as well as 80 % interval odd ratios (IORs) for the country level variables [20, 21].

We tested bivariate interactions by multiplying duration of breast feeding and maternal education, duration of breast feeding and maternal employment, immunization and maternal education, and wealth index and immunization to determine if an interaction was present.

Within countries, weights provided by the DHS for children under 5 years of age were utilized in the analysis for the level 1 data. The weights were adjusted to the survey design. Post-stratification was incorporated as a weight adjustment. The adjusted weights were used in all of the analyses. For level 2 data, between countries, weights were created and used in the analysis for each country accounting for the country’s population.

Regression analyses considering the DHS year of survey were performed to assure that the results were not biased by the different time lapses the surveys took place at each country.

Macro International provided the datasets from the 41 countries included in this study. The study was based on secondary sources without identifying information about individual participants. It was given approval by the institutional review board, Comité de Ética en Investigación, Universidad del Rosario.