Who gets into your study, how they get in, and how many you need — the decisions that set the ceiling on generalizability and power.
Define three things in writing before recruiting anyone. The target population is everyone your conclusions should apply to (“HR managers of Indian firms with 500+ employees”). The sampling frame is the concrete list or mechanism from which you can actually draw (a professional-body directory, a payroll extract) — its gaps are coverage error. The sample is who you draw and who actually responds. Every generalization claim travels back along this chain, and the chain is only as strong as its weakest link.
In probability sampling every population member has a known, non-zero chance of selection. It is what licenses statistical generalization from sample to population — confidence intervals and margins of error assume it.
| Simple random sampling | Every member has an equal chance (random number draw from the frame). Unbiased and simple, but needs a complete frame and can under-represent small subgroups by luck. |
|---|---|
| Stratified random sampling | Divide the frame into strata (sector, region, grade), sample randomly within each. Guarantees subgroup representation and improves precision; needs stratum information in advance. Weight the analysis if strata were sampled at different rates. |
| Cluster sampling | Randomly sample groups (branches, schools), then take everyone — or a random subsample — within selected clusters. Cheap for dispersed populations, but members within a cluster resemble each other, so effective sample size shrinks (design effect). |
| Systematic sampling | Every k-th unit from an ordered list after a random start. Practical for logs and registers; beware periodic patterns in the list that align with k. |
When no frame exists or resources forbid probability methods, selection is by judgment, access or referral. These samples can be entirely appropriate — purposive sampling is the correct choice for most qualitative research — but they cannot support statistical generalization, and reports must say so plainly.
| Convenience sampling | Whoever is easiest to reach (your MBA batch, mall visitors, a LinkedIn post). Fast and cheap; systematically unrepresentative in unknown ways. Describe the sample thoroughly and generalize with restraint. |
|---|---|
| Purposive (judgmental) sampling | Deliberate selection for relevance: experts, typical cases, extreme cases, maximum variation. The workhorse of qualitative and exploratory research — rigor comes from explicit selection criteria. |
| Snowball sampling | Early participants refer later ones. Often the only route to hidden or networked populations (elite informants, informal workers); over-samples well-connected people. |
| Quota sampling | Fill predetermined cells (e.g. 50 men / 50 women across age bands) by convenience within each. Mimics the look of representativeness without its statistical warrant. |
For estimation (descriptive studies), size follows precision: the margin of error you can tolerate at a chosen confidence level. Roughly, estimating a proportion within ±5 percentage points at 95% confidence needs about 385 respondents from a large population — and precision improves only with the square root of n, so halving the margin quadruples the cost.
For hypothesis testing, size follows statistical power — the probability of detecting an effect of a given size if it exists. Power is driven by four linked quantities: the significance level α (conventionally .05), the effect size you care about (the smallest effect worth detecting, not the fondest hope), sample size, and power itself (conventionally .80). Fix any three and the fourth follows; free calculators such as G*Power do the arithmetic. As orientation: detecting a medium correlation (r ≈ .30) at 80% power needs roughly 84 cases; a medium two-group difference (d = 0.5) needs about 64 per group; small effects need several hundred.