“Explain the different types of non-probability sampling techniques. Bring out the conditions of their usage with appropriate examples.” (2022)
- Non-probability sampling selects units without every element of the population having a known, non-zero chance of inclusion — selection instead rests on the researcher’s judgement, on convenience, or on the logic of an unfolding analysis.
- This is not sloppy probability sampling; it is a different logic of inference altogether, one Glaser and Strauss formalised for qualitative work as generalisation to theory rather than generalisation to a population.
- The recurring justification across all its variants is the same: no usable sampling frame exists (or can be built) for the population in question, the population is rare or hidden, or the study is exploratory and aims to discover categories rather than measure their spread.
- Five techniques dominate practice — convenience, purposive/judgmental, quota, snowball, and theoretical sampling — each defensible under a specific set of research conditions rather than as a universal second-best to random sampling.
- The thesis argued here: each technique is the correct, non-second-rate choice exactly when the population or the research goal makes probability sampling structurally impossible or theoretically inappropriate — not merely when it is cheaper or easier.
Convenience or Accidental Sampling
- Units are drawn from whoever is easiest to reach — passers-by, the researcher’s own students, respondents intercepted at a single location.
- When it is the right choice: pilot work, instrument pre-testing, or exploratory probes where the goal is to catch obvious errors in a schedule before a real study, not to describe a population.
- When it fails: used as a stand-in for a representative sample, since the selection mechanism correlates with unknown traits (availability, willingness, location), producing bias with no way to estimate its size or direction.
- It carries no claim to representativeness and should never be reported as if it does.
Purposive or Judgmental Sampling
- The researcher deliberately selects units judged, on the basis of existing knowledge, to be information-rich for the specific question at hand — key informants, atypical or extreme cases, or units chosen precisely because they test a theory’s limits.
- Karl Popper’s falsificationist logic supplies one justification for deliberately choosing untypical cases: a theory survives only if it withstands a case picked to break it, not one picked because it is easy to reach.
- When it is the right choice: elite or expert interviewing, case-study selection, and any situation where efficiency in reaching the right respondent matters more than covering a probability-defined universe — a study of custodians of an oral tradition, for instance, does not benefit from randomising who is asked.
- Its cost: entirely dependent on the researcher’s own judgement of who is “informative,” so the sample’s adequacy cannot be checked independently of that judgement.
Quota Sampling
- The population is divided into categories (sex, age, caste, income band) with a fixed quota for each, mirroring known population proportions, but the actual individuals filling each quota are left to the interviewer’s discretion in the field.
- When it is used: opinion polling and market research, where speed and known-proportion matching are valued over probabilistic rigour.
- The famous failure: the 1948 United States presidential election, where major polls using quota sampling confidently predicted a Dewey victory over Truman, who in fact won — interviewers, left free to choose whom to approach within each quota cell, gravitated toward more accessible, generally better-off respondents, introducing a systematic and entirely unmeasurable bias that the quota’s superficially correct proportions concealed.
- The flaw is not the quota structure itself but the discretion left at the final selection step — a defect probability sampling’s randomisation is specifically designed to remove.
- When it remains defensible: rapid, low-stakes estimation where a rough approximation of population proportions is acceptable and formal inference is not being claimed.
Snowball Sampling
- Existing respondents are asked to refer further respondents from within their own social network, and the sample grows outward from an initial seed or seeds — hence “snowball.”
- When it is the right, often the only, choice: hidden or stigmatised populations for whom no sampling frame exists or can ethically be constructed — sex workers, undocumented migrants, drug users, or members of a proscribed or persecuted group, where a formal list would itself be dangerous to compile.
- Its structural limitation: because the sample follows existing social ties, individuals who are social isolates — disconnected from the network through which referrals travel — are systematically invisible, so the sample may over-represent the well-networked and miss precisely the most marginal members of the population.
- Diversifying the initial seeds and tracking referral chains can partially offset this, but the underlying network-dependence cannot be fully corrected within the method itself.
Theoretical or Grounded-Theory Sampling
- Developed by Glaser and Strauss, this is sampling driven by the requirements of an emerging theory rather than by a pre-fixed plan: the next case, setting, or respondent is chosen because it promises to illuminate, extend, or challenge a category the analysis has already begun to develop.
- Data collection and analysis proceed together through constant comparison, and sampling continues until theoretical saturation — the point at which additional cases yield no new properties of the categories already identified — rather than until a predetermined sample size is reached.
- When it is the right choice: any grounded-theory study where the researcher’s goal is to generate theory inductively from data rather than test a theory specified in advance; applying probability sampling here would be a category error, since the sampling frame itself cannot be known until the theory that defines what is relevant has begun to take shape.
- Its logic of generalisation is explicitly theoretical, not statistical — the claim is that the categories generated will travel to other settings sharing the same social process, not that the sample mirrors a population’s distribution.
Comparative Snapshot
| Technique | Basis of selection | Right when | Chief risk |
|---|---|---|---|
| Convenience | Availability | Pilot/pre-test only | Unmeasurable, uncorrectable bias |
| Purposive | Researcher’s judgement of relevance | Key informants, exploratory/case work | Wholly dependent on that judgement |
| Quota | Fixed category proportions, free choice within | Rapid opinion/market estimation | Interviewer discretion within quota (1948 US election) |
| Snowball | Referral through existing networks | Hidden/stigmatised, rare populations | Social isolates invisible |
| Theoretical | What the emerging theory needs next | Grounded-theory, inductive theory-building | No statistical generalisation possible |
- “Theoretical sampling is the process of data collection for generating theory whereby the analyst jointly collects, codes, and analyses the data and decides what data to collect next and where to find them.” — Glaser and Strauss
- This formulation captures what separates theoretical sampling from every other non-probability technique on this list: sampling and analysis are not sequential stages but a single, iterative process, with the stopping rule (saturation) itself a theoretical judgement rather than a budget or a target N.
- Non-probability sampling is not a compromise forced by resource constraints; each of its variants answers a research condition that probability sampling cannot meet — an unbuildable frame, a rare or hidden population, or a theory-generating rather than population-describing purpose.
- The single recurring failure mode across the family is the same: wherever human discretion enters the selection process — the quota interviewer’s convenience, the snowball’s reliance on existing ties — an unmeasured bias enters with it, which is precisely the risk randomisation exists to eliminate.
- The corrective is not to abandon non-probability sampling but to be explicit about which condition licenses its use, and to state the resulting logic of generalisation — theoretical, not statistical — honestly rather than dressing a purposive or snowball sample in the language of representativeness.
- Contemporary digital research on hard-to-reach populations (online communities of stigmatised or minority groups, for instance) has extended snowball logic into networked, platform-mediated recruitment, but the same isolate-invisibility problem persists in a new form.
- The enduring methodological lesson is that the choice of sampling technique should follow from the nature of the population and the purpose of the inquiry, not from habit — treating every non-probability sample as a lesser substitute for a probability sample misunderstands what each is actually for.