Sampling and representativeness

Sampling connects a research question to the people, places, or records that can answer it. Representativeness is a claim about a defined population and selection process. A large number of interviews, an even gender balance, or coverage of several villages does not establish that claim by itself. Social impact assessment commonly needs both population estimates and detailed accounts of particular experiences; design each component for its purpose.

Define what will be estimated

Specify the target population, observation period, eligibility rules, and unit: people, households, workers, businesses, parcels, or resource-use episodes. Identify the groups for which separate estimates are required. A sample adequate for an overall mean may provide little precision for a small livelihood group or an individual village.

Distinguish an affected-person register from a socioeconomic survey. A register supports identification and case follow-up; a sampled survey estimates specified characteristics within its coverage. Neither purpose makes the records interchangeable. The questionnaire guide explains how respondent selection and linked observation units affect measurement.

Use probability selection for design-based estimates

Construct or assess the sampling frame, documenting omissions, duplicate entries, and outdated units. In a probability design, eligible units within the defined survey population have known, nonzero selection probabilities. Preserve selection at every stage, including selection of locations, households, and individuals where applicable. Random selection of villages followed by convenient household recruitment does not preserve a probability household design.

The UN's Designing Household Survey Samples, chapters 3, 6, and 7, connects selection probabilities with weights and sampling-error estimation. Keep the selected-unit list, strata, clusters, field outcomes, and design changes as part of the evidence record.

Plan sample size for an indicator and design

For one proportion under a simple random approximation, a common initial calculation is n₀ = z² p(1 − p) / e². Here e is an absolute margin of error, p is the anticipated proportion, and z corresponds to the confidence level. With p = 0.5, e = 0.05, and z = 1.96, the calculation gives 384.16, rounded up to 385 completed observations before other design considerations. Penn State STAT 506, lesson 2

Clustering, required reporting domains, nonresponse, and the chosen estimator change the design requirements. A finite-population correction must match the sampling stages. Power calculations for detecting change or differences require an effect size and other assumptions; they answer a different question from the precision of one descriptive estimate. The detailed baseline sampling guide develops these calculations and their limits.

Preserve weights and uncertainty

Start design weights from inverse selection probabilities. Document subsequent nonresponse or calibration adjustments and the information and assumptions they use. Weighting does not automatically repair omitted groups or unmeasured differences between respondents and nonrespondents.

Use an analysis procedure that incorporates the relevant strata, clusters, and weights. Report the population, denominator, unweighted observations, estimate, and uncertainty together. A 95% confidence procedure has its coverage interpretation over repeated samples under its assumptions; it does not cover every source of measurement or selection error. Penn State STAT 500, lesson 5

Use purposive sampling for qualitative questions

Select participants for relevant knowledge and contrasting experiences, then document recruitment and adapt as the inquiry develops. Seek cases that challenge an emerging explanation. Record whether further interviews add new meanings, conditions, or perspectives, including differences between groups. A statement that saturation was reached needs an explanation of what was assessed and how.

Guest, Bunce, and Johnson (2006) remains an important empirical reading on interview saturation. Its results should be interpreted within the study's setting and analytic approach. Fugard and Potts (2015) introduced a planning model for detecting theme instances; Fugard's later clarification explains the role of relevant recruitment and assumptions about theme occurrence. Such calculations support a particular planning rationale, rather than establish a universal interview target or population prevalence.

Document the achieved study

Describe exclusions, refusals, inaccessible locations, deviations, and remaining gaps. Keep qualitative findings and probability estimates identifiable when combining them in an assessment. The historical Maluku sampling case illustrates a proposed field design; its planning quantities require appraisal before reuse.

The methods references preserve the earlier qualitative, statistical, and software reading trail. One original NIST reading concerns normal-distribution tolerance intervals, which concern coverage of population values and serve a different purpose from confidence intervals for a population parameter.