Nth name selection: Generally used for test panel record selection. Every nth record is chosen. If you are selecting, say, 10,000 records out of 1 million, every 100th record is chosen. (In this case, n = 100.) This technique can cause problems if there is an intrinsic pattern within the data set. For example, if you want a 50% sample, and the data set is ordered male, female, male, female, you could wind up with all male names, although few lists are like this. Or in list order fulfillment, 10 orders of a 10,000 nth selection could result in the same 10,000 names being selected 10 times.
Nthing with a seed: Often used in list order fulfillment to ensure against the problems inherent in normal nth name selection. For example, if 10 catalog companies ordered list tests of 10,000 from a universe of 100,000, this process would prevent all of the companies from receiving the same records. For each company, one record from the first 10 (i.e., the universe divided by the desired panel size) would randomly be selected, and then every 10th record subsequent to it.
True random sampling: Each record is selected randomly and has an equal chance of selection. For instance, to select 10,000 out of a 100,000 base, the first name chosen may be #682, then, #99,204, then, #88, and so on.
Stratified sampling: A process to improve the representativeness of test panels compared with normal sampling. This is done by dividing the universe into subgroups based on a factor that correlates with what you are trying to measure, and then sampling within these subgroups.
For example: If you were trying to measure the next 12-month sales, you would: 1) rank the universe based on previous 12-month sales, 2) divide into equal 25 panels, and 3) nth-select within each panel. This technique is often used when dealing with extreme price points.
Test panel design Holdout panels: Known customers who receive no database-driven promotions. These customers can be influenced only by other, non-database driven media.
Longitudinal test: A long-term test (often 6-12 months) using an A/B split to measure cumulative behavior over time. Multiple panels can be used. Panel A receives multiple database promotions; Panel B receives no database promotions. This design is useful because database promotions often change behavior slowly, which cannot be measured with a single test.–CBW
Mean: The average result. (In the illustration on p. 66, the mean = 1.0%.)
Median: If all of the observations in an analysis were listed from high to low, the median is the one in the middle, regardless of value. For instance, if you took a sample of 101 observations, the median would be the actual value of the 50th observation.
Mode: The observation value that happens most frequently.