Introduction to Strata and Clusters

Strata are internally homogeneous but externally heterogeneous, meaning items within a stratum are similar while different strata differ from each other. In contrast, clusters are internally heterogeneous but externally homogeneous; members of a cluster vary, yet clusters resemble one another. When stratifying, each stratum must be sampled to preserve representativeness, and a person can belong to only one stratum.

Systematic and Stratified Sampling Rationale

Systematic sampling relies on a sorting variable that should be strongly correlated with the outcome of interest. The same principle applies to stratification: the stratifying variable must be highly correlated with the outcome. Using an irrelevant variable—such as favorite color in a smoking study—fails to improve precision. Stratification can also ensure representation of specific subgroups, for example minority groups in a smoking prevalence survey.

Implications of Stratification

Stratification increases precision because each stratum is treated as a separate population, but this precision comes at a dramatically higher cost. Adding multiple stratifying variables (e.g., gender, race, population density) can generate dozens of strata; a design with 24 strata, a design effect of 2, and a 50 % response rate required about 36,879 participants. Incentives of $30 per participant and recruitment fees of $120 per participant alone total over $5.5 million, and the full study can easily exceed tens of millions of dollars.

Cluster Sampling Rationale and Characteristics

Clusters are often groups of people living close together—households, schools, or neighborhoods. They can be defined arbitrarily and do not need to be naturally occurring. The primary advantage of cluster sampling is cost and time efficiency: “Clustering” (physical proximity) leads to “clustering” (statistical similarity). Real‑world clusters vary in size and composition, which can introduce bias if not managed. Strategies to address variability include splitting large clusters, combining small ones, stratifying clusters by size, or using Probability Proportional to Size (PPS) sampling.

Sampling Techniques and Their Impact

• Simple Random Sampling (SRS) draws individuals directly from the entire frame, providing the lowest variance.
• Single‑Stage Cluster Sampling selects whole clusters at random and includes every member, saving time but increasing variance.
• Two‑Stage PPS Cluster Sampling first selects clusters with probability proportional to their size, then samples a fixed number of individuals within each selected cluster.
• Stratified Two‑Stage Cluster Sampling adds a stratification layer, sampling clusters within each stratum before selecting individuals.

Across these designs, cluster‑based methods generally produce larger variance and wider confidence intervals than SRS.

Confidence Intervals and Variance

A confidence interval gives a range where the true population parameter is likely to fall. A 95 % confidence interval means that if the study were repeated an infinite number of times, 95 % of those intervals would contain the parameter. Variance directly influences interval width: higher variance yields larger confidence intervals. Because cluster sampling inflates variance relative to SRS, it also expands confidence intervals, reducing the precision of estimates.

Practical Exercise Using the K‑qu Platform

Using the K‑qu software, several sampling designs were applied to estimate the mean of Metabolite X:

1. SRS on the combined frame produced the smallest variance.
2. Single‑Stage Cluster Sampling selected clusters randomly and included all members, resulting in higher variance.
3. Systematic Sampling sorted clusters by population size before selection, also showing increased variance.
4. Two‑Stage PPS Cluster Sampling chose clusters proportionally to size and then sampled individuals, further widening confidence intervals.
5. Stratified Two‑Stage Cluster Sampling stratified clusters, then sampled within strata, delivering the greatest variance among the methods tested.

These results illustrate how sampling design choices affect both the point estimate and its associated uncertainty.