Description
Test whether a yes/no outcome clusters within teams, households or accounts more than equal individual risk would produce, then model the clustering.
Data come as groups of different sizes, each with some number of affected members: users per account who hit a bug, employees per team who adopt a tool. Under the null hypothesis every individual has the same risk and individuals are allocated to groups at random, which fixes expected counts of groups with 0, 1, 2 and more affected members. Conditioning on group sizes and the total affected, the exact p-value sums the probabilities of all tables no more likely than the one observed, computed by enumeration or simulation. Standardised residuals, shown as a bubble plot, locate the groups driving any departure. Zelterman stresses that clustering (cases concentrated in few groups) differs from high incidence, and that groups found through an affected member must have that member removed. If clustering is present, beta-binomial, Altham or weighted binomial models describe it and are compared by deviance.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Evaluation, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Tabulate, for each group size, how many groups have 0, 1, 2 and more affected members.
- If groups were recruited because one member was affected, remove that index member from each group.
- Optionally test overall incidence against a reference rate first, then test clustering separately.
- Compute expected counts under equal individual risk (hypergeometric within each group size, or binomial with p equal to total affected over total individuals).
- Obtain the exact p-value by enumerating tables with the same group sizes and total affected, or by Monte Carlo simulation when enumeration is too large.
- Inspect standardised residuals in a bubble plot to find the clusters.
- If clustering is present, fit beta-binomial, Altham or weighted binomial models, compare them by differences in deviance, and add covariates where available.
Sources
- Discrete Distributions: Applications in the Health Sciences (Zelterman, 2004)
Also known as
- Exact test of homogeneous risk across groups
- Clustering versus incidence test
- Family frequency distribution test
- Altham and weighted binomial clustering models
- Exact test of homogeneous risk across groups of unequal size (clustering of binary events within families, households or teams), with dependent-Bernoulli follow-up models
Categories
- Analytics & Behaviour
- Statistics & Modelling