Solar Drinking Water Disinfection (SODIS) to Reduce Childhood Diarrhea - Recent Advances and Issues in Environmental Science - page 193

the latter method requires a larger number of clusters to produce consistent estimates
[26].
The crude (unadjusted) model included only the design factors and the interven-
tion effect [12, 27]. Further models included potential confounders (selected a priori:
child's age, sex, child hand-washing behavior, and water treatment at baseline). Fol-
lowing an evaluation of the best fi t, the GLMM included the log link function for
negative binomial data (IR) and logit for binomial data (PR and SD). Denoting the link
function of the outcome Y by g ( E ( Y )), the crude and adjusted models were: g ( E ( Y ijk ))
= μ+ B i +τ j +ξ ij , and g ( E ( Y ijk )) = μ+ B i +τ j +ξ ij + x'b where Y ijk denotes the observed outcome
value for the k th individual from a community allocated to the j th intervention, in the
i th pair, μ is the general mean, B i is the random effect of the i th pair ≈N (0, σ 2 p ), τ j is the
fi xed effect of the SODIS intervention, and ξ ij is the random effect of the interaction of
the i th pair with the j th intervention applied to the community ≈N (0, σ 2 pt ) (signifying
the within-pair cluster variance and used as error term for τ j ), x is the vector of poten-
tial confounding factors, and b the vector of the corresponding regression coeffi cients.
The intracluster correlation coeffi cient (ICC) and the coeffi cient of between-clus-
ter variation ( k ) were calculated after data collection to validate the degree of clus-
tering and our assumptions for the sample size. The ICC and k were estimated from
the unscaled variance of the IR's GLMM. To estimate the uncertainty of ICC and k ,
we obtained the 95% credible region (Bayesian equivalent of 95% CI) through an
analogous Bayesian hierarchical regression [28]. Noninformative priors were used.
The statistical analyses were performed using SAS software v9.1 (PROC GLIMMIX,
SAS Institute Inc.) and WinBUGS v1.4 (Imperial College and MRC).
Participant Flow and Recruitment
Among the 1,187 households in the 22 communities there were 546 that met the in-
clusion criteria (Figure 1). The median number of participating households with chil-
dren <5 year per community was 22. Because of political unrest and national election
campaigns in 2005 a period of 6 month passed between the baseline and the start of
follow-up. Subsequently, 62 households (102 children) were no longer traceable be-
fore randomization, and 59 households (37 intervention, 22 control) were lost before
data collection had started. The loss to follow-up was balanced in intervention and
control arms. Data were obtained from 376 children (225 households) in the interven-
tion and 349 children (200 households) in the control arm, thus reaching our originally
planned sample size.
Follow-up started in June, 2005 and ended in June, 2006. During the 51 week
of the study, information on the occurrence of diarrhea was collected for 166,971
person-days representing 79.9% and 78.9% of the total possible person-days of child
observation in intervention and control arms. We excluded from the potential observa-
tion time the experience of 94 children who dropped out before the start of follow-up.
National festivities, holidays, and political unrest over the entire year amounted to
further 9 week during which outcome surveillance needed to be suspended. The main
reasons for incomplete data collection were migration (28%) and withdrawal (67%).
Supervisors reevaluated the outcome during 984 unannounced random home visits,