Description
Plan how many people to screen to fill a quota in each of two categories with stated confidence, and stop early if a category proves too rare.
Recruiters often need at least c people in each of two groups (users and non-users of a feature, left- and right-handed participants) and screen people one at a time. The number screened until both quotas are filled follows the maximum negative binomial distribution, the larger of two dependent waiting times; its upper tail gives the number of screens that fills both quotas with, say, 95 percent probability. With an even split and 7 needed per group, about 17 screens are expected and 22 give 95 percent assurance. The same distribution supplies a sequential test: if a critical number is reached without filling both quotas, conclude the category proportion is extreme (near 0 or 1) rather than moderate. A without-replacement version applies to small known pools such as a customer panel. Recording the split when screening stops allows the proportion to be estimated.
Attributes
- Discipline
- Research Design & Rigor
- Loss type
- Unmigrated
- Goal
- Support a decision, Prioritise
- Stage
- Discovery, Definition
- Data
- Analytical / expert
- Effort
- Hours
Steps
- Fix the quota c needed in each of the two categories.
- Set a planning value for the proportion p in the first category, or a plausible range (for example 0.1 to 0.9) to serve as the null hypothesis for a test.
- From the maximum negative binomial upper tail (or the negative hypergeometric version for a small finite pool), read the number of screens that fills both quotas with the desired probability.
- For a stopping test, take the critical number at the boundary of the null range and alpha level; screen sequentially and stop when both quotas are filled or the critical number is reached.
- If the critical number is reached first, conclude the category is rarer (or commoner) than assumed and revise the recruitment strategy.
- Record how many of each category had been seen when screening stopped, so p can be estimated without ambiguity.
Sources
- Discrete Distributions: Applications in the Health Sciences (Zelterman, 2004)
Also known as
- Maximum negative binomial distribution
- Maximum negative hypergeometric distribution
- Frequency quotas
- Noah's ark problem
- Screen until both types are seen
- Quota-screening plan and stopping test: trials needed to observe at least c of each of two categories
Categories
- Research Operations
- Research Design & Rigor