3-year phD position in medical statistics in Tours (France).
If you are interested, please send an email to Bruno Giraudeau ([log in to unmask]) for more information
PhD position: Clustering effect measurement for binary outcomes
Supervisor: GIRAUDEAU Bruno ([log in to unmask])
Co-supervisor: ELDRIDGE Sandra ([log in to unmask])
Research unit: INSERM U1246 (SPHERE)
Grant: INSERM – Val de Loire Region – 3 year position (Oct 2017 – Sept 2020)
Context: A cluster randomized trial is a trial in which intact social units (named clusters) are randomized rather than individuals (1). In such a situation, responses from individuals belonging to a common cluster are more similar than those from individuals who belong to different clusters. This situation is named the clustering effect. It is usually quantified by the intraclass correlation coefficient (ICC) (2). The coefficient is defined as the correlation between two observations from a common cluster (whatever the observations and whatever the cluster) or the ratio of the between-cluster variance to the overall variance. Estimating ICCs is part of the statistical analysis of a cluster randomized trial and helps in interpreting the trial results.
The clustering effect is not limited to the cluster randomized trial context. Indeed, when one is interested in center effects, this also comes down to study the clustering effect. Otherwise, it makes also sense when one deals with intrafamilial correlation, which is the context in which ICCs were first considered.
When data are binary, the ICC depends on the prevalence (it increases as the prevalence increases up to 50% and then decreases as the prevalence is close to 100%) (3), which hampers an easy interpretation of its values. Indeed, when the assessed intervention shows efficacy, success rates differ between groups, which implies that ICC estimates are different between groups, as an artefact. Other coefficients have been proposed to quantify the clustering effect: the coefficient of variation (4); the Kappa coefficient, although it also depends on the prevalence (2); or the R coefficient (5). Otherwise, if we hypothesize that the binary outcome is actually the result of a binarisation of a continuous outcome, the associated ICCs have been found to be related (6). This relationship is still true in the framework of cluster randomized trials (7), and perhaps switching from one ICC to the other one would allow for quantify the clustering effect without the artefact due to prevalence.
Objective: Comparing the different approaches used to quantify the clustering effect for binary outcomes.
Methods: First, we will attempt to identify all the parameters that have been used to quantify the clustering effect. For this, we will use keywords (considering both the clustering effect as defined in the cluster randomized context and also in the center-effect context) and also a snow-ball approach. Then, we will compare parameters by using Monte-Carlo simulation studies to assess the statistical properties of the different parameters. Results will be applied to real data available in the lab.
Output: Results will allow for identifying the best method(s) to quantify the clustering effect in cases of binary data. This work will translate into recommendations for trialists who conduct cluster randomized trials.
1. Donner A. Design and analysis of cluster randomization trials in health research. London: Arnold; 2000.
2. Eldridge SM, Ukoumunne OC, Carlin JB. The Intra-Cluster Correlation Coefficient in Cluster Randomized Trials: A Review of Definitions. Int Stat Rev. 2009;77(3):378 94.
3. Gulliford MC, Adams G, Ukoumunne OC, Latinovic R, Chinn S, Campbell MJ. Intraclass correlation coefficient and outcome prevalence are associated in clustered binary data. J Clin Epidemiol. mars 2005;58(3):246 51.
4. Hayes R. Cluster randomised trials. Boca Raton: CRC Press; 2009.
5. Crespi CM, Wong WK, Wu S. A new dependence parameter approach to improve the design of cluster randomized trials with binary outcomes. Clin Trials Lond Engl. déc 2011;8(6):687 98.
6. Kirk D. On the numerical approximation of the bivariate normal (tetrachoric) correlation coefficient. Psychometrika. 1973;38:259 68.
7. Caille A, Leyrat C, Giraudeau B. Dichotomizing a continuous outcome in cluster randomized trials: impact on power. Stat Med. 30 oct 2012;31(24):2822 32.
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|