we know cluster sampling is done for large population with in naturally occurring clusters . It is mostly used in evaluation of health services . In India for assessment of UIP ( Universal immunization program) , we take 30 clusters with 7 sample from each cluster. Similarly for NIDDCP( National iodine deficiency control program) we take 30(cluster)x 90(children) based on School enrolment ratio with equal number of boys and girls. How we ascertain the number of cluster and sample to be taken in each cluster?
There is no standardization yet for cluster sizing. You may fix it according to the behaviour of your model/algorithm/method. However, the best way is to split the clusters/classes in equal training & testing samples. For example, if you want to perform parmetric methods for classification, the best way is to use 10 to 20% for traiing and the remaining for testing via few sample trials. In non-parametric the suggested way is to use 50% for training and 50% for testing...
I think you are asking about statistical power calculations for clustered studies. There are several ways to do this and it can get quite complicated, but I'd start by reading about intra-class correlations in clustered studies and the design effect.
e.g., Article What Is an Intracluster Correlation Coefficient? Crucial Con...
Dear Suraj Kapoor ,
The optimal sample size per cluster only depends on the cluster-to-person cost ratio c/s and on the ICC and is between 7 and 70 if the cost ratio is between 5 and 50 and the ICC is between 0.01 and 0.10. The number of clusters, and thereby also the sampling budget, additionally depends on the effect size and power.
Article Calculating sample sizes for cluster randomized trials: We c...
Regards,
Shafagat
- Badges
- Science topic
- Similar topics
- Philosophy
- Philosophy Of Science
- Reasoning