This report utilizes secondary data from the Nigerian Demographic and Health Survey [9]. Demographic and health surveys are an international series of nationally representative surveys conducted in middle and lower income countries [10].
The study population in this analysis was the 25,004 under five children extracted from the records of 104,808 respondents who responded to the household questionnaire. The outcome measure in this analysis was occurrence of fever in children less than five years of age in the last two weeks preceding the survey. The explanatory variables were sex of child, age of child, whether or not household possess a bed net, whether child slept under bed net night before survey, type of bed net (treated or untreated), type of place of residence (rural/urban) and wealth index; a measure of poverty. This index serves as a proxy for measuring the long-term standard of living. It was based on data from the household's ownership of consumer goods; dwelling characteristics; type of drinking water source; toilet facilities and other characteristics that are related to a household socio-economic status such as car, radio, television, gas cooker, iron, boat/ship, telephone, source of water, electricity supplies, and type of wall, floor and roof, etc. To construct the index, each of these assets was assigned a weight (factor score) generated through principal component analysis, and the resulting asset scores were standardized in relation to a standard normal distribution with a mean of zero and standard deviation of one [11]. Each household was then assigned a score for each asset, and the scores were summed for each household. Individuals were ranked according to the total score of the household in which they resided. The sample was then divided into quintiles from one (lowest) to five (highest). A single asset index was developed on the basis of data from the entire country sample and this index was used in the analysis. The data were analyzed using Stata 10.0 [12].
Frequency tables were generated for relevant variables. Descriptive statistics such as means, medians and standard deviations were used to summarize continuous variables. Relationships between the outcome measure (occurrence of fever) and each explanatory variable was first explored using the chi squared test in bivariable analysis before a multilevel logistic regression was carried out. Explanatory variables found to be associated with the outcome were included in the logistic model based on the magnitude of the chi squared values. Odd ratios and 95% CIs were computed. A random effect logistic model was fitted that included fixed effects and group-level intercepts as random effects [13]. This allowed for a two-level, inherently nested nature of the data: individuals (children) nested within clusters/regions. Nigeria has six geopolitical zones and there is variation in all characteristics including fever patterns in all the zones.
Multi-level models allow the additional information provided by knowing which cluster a child comes from to be taken into account in modeling the relationship between independent and dependent variables. This model was run using the GLLAMM routine in Stata. In using this approach, the correlation between children from the same cluster or region arises from their sharing specific but unobserved properties of the respective regions. A random effect logistic regression model for the data is given by
Where b1, b2, b3 and b4 represents fixed effects and Ui represents a random effect. Ui is a random variable with a mean of zero and constant variance. Ui, is the estimation of the variance across all of the regions involved in the study. If the variance is large then the outcome of interest is dependent on the region, if the variance is small then the variations in outcome of interest may be explained by the measured characteristics alone.
Given Ui, the responses from the same regions are mutually independent that is the correlation between children from the same regions is completely explained by them having been observed in the same regions. The variance τ2 measures the degree of heterogeneity in the probability of experiencing fever that cannot be explained by the classification into the 2 categories (fever: yes Vs no)
An important measure that describes these dependences in the data is called the intra-class correlation coefficient (ICC); this statistic measures the extent to which individuals within the same group are more similar to each other than they are to individuals in different groups. The intraclass correlation (ICC), ρ was calculated using:
Where τ = estimated variance and π = 3.142
The Gllamm approach
This is a stata procedure for fitting generalized linear models, a class of regression models for univariate responses with density from an exponential family [14]. The logistic regression model was specified from the family of generalized linear model with the logit link and binomial distribution using the link ( ) and family ( ) options. The option i(region) specifies the desired within region correlation matrix. The eform produces the estimated odds ratios. The full syntax is given below:
xi: gllamm fever i.place of residence i.sex of child i.childsage i.wealthnw, i(region ) family(binomial) link(logit) adapt nip(5), eform