The analyses of both prevalence of infection and incidence of clinical episodes were based on a generalised estimating equation (GEE) approach that allowed accurate modelling of both the variation in the outcomes and the correlations between repeated measurements in individual children.
The prevalence of infection at the 8 or 9 weekly bleed time points were analysed using logistic GEE models via the XTLOGIT procedure in STATA 8.0. Preliminary analyses showed that the correlations between measurements in individual children were best modelled using an exchangeable correlation structure. A semi-robust Huber/White/sandwich estimator of variance was used to assure valid standard errors. Best fitting models were determined by backward elimination using Wald's Chi-square tests for individual variables. The analyses of age and spatial trend and predictors of infection were based exclusively on data from double bleed time points (i.e. weeks 9–60) using the combined diagnostic results from both samples (see above). As only one sample was collected at baseline and final bleeds (i.e. weeks 0 and 69), only prevalence data from the first sample collected at double bleed time points was used for the analyses of time trends across the entire time of follow-up. Difference in log-transformed parasite densities among age groups were analysed using a normal GEE model with exchangeable correlation structure and semi-robust variance estimator.
For estimation of incidence of clinical episodes in each 8 or 9 week follow-up interval, clinical malaria episodes were defined as a fever (i.e. axillary temperature >37.5°C) or history of fever during the last 48 hrs in the presence of a light microscopically detectable parasitaemia observed in either passive or active follow-up. Febrile episodes with only LDR-FMA detectable parasitaemia were not considered to be malarial episodes. A density cut-off value for clinical disease was set based on a pyrogenic threshold of 2500 parasites/µl for P. falciparum and 500 parasites/µl for all other species [32] . Moderate and high density episodes were defined as febrile episodes with parasitaemias exceeding the pyrogenic threshold 4 and 20-fold, respectively. For each interval, children were considered at risk from the 1st day after the 2nd or only blood sample was taken. Cross-sectional bleed days were thus considered part of the preceding interval and clinical episodes detected during the cross-sectional bleeds were included in that interval. Children were not considered at risk for 2 weeks after treatment with Coartem® and 4 weeks after treatment with AQ plus SP.
As preliminary analyses showed significant overdispersion in the number of episodes per child, a negative binomial model GEE model (based on XTNBREG procedure) was used for the analyses of incidence rates. As in the other GEE models, an exchangeable correlation structure and semi-robust variance estimator were used and best fitting models determined by backward elimination using Wald's Chi-square tests for individual variables. Differences in participant characteristics at enrolment were assessed using Chi-square and Fisher's exact tests. All analyses were done using the STATA 8 statistical software package (College Station, TX).
The prevalence of infection at the 8 or 9 weekly bleed time points were analysed using logistic GEE models via the XTLOGIT procedure in STATA 8.0. Preliminary analyses showed that the correlations between measurements in individual children were best modelled using an exchangeable correlation structure. A semi-robust Huber/White/sandwich estimator of variance was used to assure valid standard errors. Best fitting models were determined by backward elimination using Wald's Chi-square tests for individual variables. The analyses of age and spatial trend and predictors of infection were based exclusively on data from double bleed time points (i.e. weeks 9–60) using the combined diagnostic results from both samples (see above). As only one sample was collected at baseline and final bleeds (i.e. weeks 0 and 69), only prevalence data from the first sample collected at double bleed time points was used for the analyses of time trends across the entire time of follow-up. Difference in log-transformed parasite densities among age groups were analysed using a normal GEE model with exchangeable correlation structure and semi-robust variance estimator.
For estimation of incidence of clinical episodes in each 8 or 9 week follow-up interval, clinical malaria episodes were defined as a fever (i.e. axillary temperature >37.5°C) or history of fever during the last 48 hrs in the presence of a light microscopically detectable parasitaemia observed in either passive or active follow-up. Febrile episodes with only LDR-FMA detectable parasitaemia were not considered to be malarial episodes. A density cut-off value for clinical disease was set based on a pyrogenic threshold of 2500 parasites/µl for P. falciparum and 500 parasites/µl for all other species [32] . Moderate and high density episodes were defined as febrile episodes with parasitaemias exceeding the pyrogenic threshold 4 and 20-fold, respectively. For each interval, children were considered at risk from the 1st day after the 2nd or only blood sample was taken. Cross-sectional bleed days were thus considered part of the preceding interval and clinical episodes detected during the cross-sectional bleeds were included in that interval. Children were not considered at risk for 2 weeks after treatment with Coartem® and 4 weeks after treatment with AQ plus SP.
As preliminary analyses showed significant overdispersion in the number of episodes per child, a negative binomial model GEE model (based on XTNBREG procedure) was used for the analyses of incidence rates. As in the other GEE models, an exchangeable correlation structure and semi-robust variance estimator were used and best fitting models determined by backward elimination using Wald's Chi-square tests for individual variables. Differences in participant characteristics at enrolment were assessed using Chi-square and Fisher's exact tests. All analyses were done using the STATA 8 statistical software package (College Station, TX).
Free full text: Click here
