Three steps a descriptive survey needs before fieldwork. How many interviews give the precision you want once clustering and non-response are allowed for. How to divide them between strata. Which clusters to visit, drawn with probability proportional to size from your own list, with the weight each interview will carry. Every number is shown with its working.
Purpose
Precision
0.5 gives the largest sample. Use it when there is no earlier estimate.
Design
Sample size
What changes it
Method
For one proportion the simple random sample is n₀ = z²p(1−p)/e², and for a mean n₀ = z²σ²/e² (Cochran 1977, chapter 4). With a known population N the finite population correction gives n₀/(1 + (n₀−1)/N). The design effect 1 + (m−1)ρ multiplies that (Kish 1965), and the result is divided by the response rate. For a change between two independent rounds, each round needs (z₁₋α/₂ + z₁₋β)²[p₁(1−p₁) + p₂(1−p₂)]/(p₂−p₁)², or 2σ²(z₁₋α/₂ + z₁₋β)²/δ² for means, times the design effect. Sizes are rounded up.
Strata
The standard deviation is optional. For a proportion p it is √(p(1−p)): 0.50 at p = 0.5, 0.40 at p = 0.2. Without it Neyman allocation cannot be computed.
Allocation
Press Allocate.
Method
Proportional allocation gives each stratum n·Nh/N. Neyman allocation gives n·NhSh/ΣNhSh, which minimises the variance of the overall mean for a fixed total (Neyman 1934; Cochran 1977, section 5.5). Equal allocation gives n/H, and suits a survey whose first purpose is a separate estimate for each stratum. A stratum never gets more than its population, and the excess is shared among the others. The standard error shown is √Σ Wh²Sh²(1 − nh/Nh)/nh for a simple random sample within each stratum, before any design effect.
Frame
The example is invented: 60 villages in four districts. The file is read in this page and is not uploaded.
Selected clusters
Load a frame to begin.
Method
Systematic sampling with probability proportional to size: the sizes are cumulated in file order, the interval is the total divided by the number of clusters, a random start is taken within the first interval, and a cluster is selected each time the running point falls inside its range (Cochran 1977, section 9A.9). A cluster at least as large as the interval is taken with certainty and the interval is recomputed for the rest. With strata, clusters are divided between them in proportion to their size and each stratum has its own start. A cluster of size Mi is selected with probability k·Mi/M, and a household within it with probability take/Mi, so the weight is the inverse of the product. Where no cluster is smaller than the take or selected with certainty, every household has the same weight.