2. Analysis of cluster‐RCTs reporting clinical outcomes.
Trial ID | Unit | Mean cluster population | Number of clusters | Cluster adjustment by trial authors | Approximate ICC calculated by review authors1 | Cluster adjustment by review authors2 |
Costa 2007 BRA | Geographical area | 11 | 34 | "We specify a model that explicitly considered the effect of aggregation of the individual in clusters (cluster effect) and used methods of robust estimation of variance. Data analysis was performed using STATA software." | Unable to calculate because the raw data were not presented. | None necessary. |
Emami 2009 IRN | Urban sectors | 635 | 12 | None (analysed at the individual level). | ‐ | SE adjusted for clustering using the ICC from Rojas 2006 COL. |
Kroeger 2002 VEN | City sectors | 210 | 14 | 'We compared data using a paired t test, weighting the data according to the sector size. We also used Wilcoxon's matched pairs test because the small number of pairs made it difficult to assess whether the underlying distribution of the differences was normal'. | Unable to calculate as authors only presented mean difference adjusted for clustering. | RR was calculated from raw data and the SE adjusted for clustering using the ICC from Rojas 2006 COL. |
Picado 2010a ASIA | Hamlets | 761 | 26 | "Adjusted analyses were carried out in two stages...a standard individual level logistic regression model to calculate expected number of events for each cluster ignoring the intervention...The adjusted intervention effect was calculated with these residuals in a paired t test". | 0.0010 | None necessary. |
Reyburn 2000 AFG | Household | 5 | 957 | "Because the interventions were allocated at household level, the data were analysed by a random effects logistic regression model to adjust for the possibility that individuals within a household might be more similar with respect to the intervention outcome than individuals from other households". | 0.0321 | Converted from OR to RR using the formula: RR = OR/(1‐ACRx(1‐OR)). |
Rojas 2006 COL | Village | 182 | 20 | "Once the final model was defined, the generalized estimating equations method was used to estimate the parameters while taking into account the correlation of observations within villages". | 0.0034 | None necessary. |
Werneck 2014 BRA | City blocks containing ≈ 60 households | 70 | 40 | "using Poisson population‐average models from generalized estimating equations with robust variance, an exchangeable correlation model, and designating each block as the clustering level". | ‐ | None necessary. |
Abbreviations: BRA = Brazil; IRN = Iran; VEN = Venezuela; AFG = Afghanistan; COL= Colombia; ICC = intra‐cluster correlation co‐efficient; SE = standard error; RR = risk ratio; OR = odds ratio. 1We calculated the ICC by comparing the cluster‐adjusted SE with the unadjusted SE to calculate the design effect (DE) and then using the formula: DE = 1+(M‐1)*ICC where M=mean cluster size. 2We chose the ICC value by looking for the trial with the most similar size of clusters and number of clusters.