Design effect
A design effect compares an estimator's variance under a sampling design with its variance under a specified simple-random-sampling reference.
A design effect belongs to an estimator, outcome and reference design, not to a dataset alone. Variance inflation by D corresponds to a standard-error multiplier sqrt(D), not D. [1]
For the equally weighted mean of N observations in fixed independent groups of sizes m_g, assume a common marginal variance and common within-group pair correlation rho. Summing covariances gives D = 1 + rho * (sum(m_g^2)/N - 1) relative to N independent observations with that variance. Equal sizes m reduce this to 1 + (m - 1) * rho. The unequal-size expression weights large groups more heavily. [2] [3]
Here N/D is a variance-equivalent sample size for this mean. It is not a count of independent groups, a degree of freedom or permission to substitute N/D into an exact binomial interval. Other weights, unequal risks, censoring or dependence across groups need a different model. Rack membership alone does not establish independence. [3]
Relations
Citations
- Survey Methodology (2020): Considering interviewer and design effects when planning sample sizes, section 1
- Cameron and Miller (2015): A Practitioner's Guide to Cluster-Robust Inference, sections II.B and VI
- Hesela: fixed-group covariance calculation, estimator scope and unequal group sizes
Sources & caveats
- The storage example is an exact calculation under specified covariance assumptions, not a measured intraclass correlation or a survey-design estimate for a real fleet.