Sampling and Fieldwork with Hard-to-Reach Populations

A sampling frame is the concrete list a researcher draws from: an electoral roll, a hospital patient register, a company payroll. Survey research assumes such a list exists, is reasonably complete, and can be sampled at random. Criminology routinely works with populations for which no such list exists at all. There is no register of active burglars, no directory of people who deal drugs, and no membership roll for a street gang that a researcher could obtain and sample from.

Official records create only a partial and distorted substitute for a frame. Arrest data, prison admissions and conviction records capture the subset of offenders who were caught and processed, which is not a random subset of everyone who offends.

In India, for instance, the annual crime statistics compiled by the National Crime Records Bureau record only offences that were reported to police and formally registered, so anyone who wants to study offending that never reaches that stage has no frame to sample from at all.

Studies built only from official records systematically describe the least careful, least resourced or unluckiest offenders, and say very little about the larger group who were never caught. This gap between the offending population and the officially processed population is sometimes called the dark figure of crime, and it is precisely why fieldwork-based sampling exists as a distinct methodological tradition within criminology.

Researchers studying corporate or financial offending face a related but different version of the problem. Companies and their records are relatively easy to enumerate, so a study of organised, white-collar or cybercrime can sometimes build a frame from regulatory filings or industry directories, then sample firms or executives from it.

The frame exists, but access to honest answers about wrongdoing inside it is the harder problem, which shows that the missing-frame problem and the access problem, discussed later in this topic, are related but separable challenges.

Because no frame exists for most offender populations, researchers cannot claim their final sample was drawn with a known, equal probability from a defined universe. This has a direct methodological consequence: classical statistical inference, which assumes every population member had a calculable chance of selection, does not apply cleanly to snowball or purposive samples of hidden populations. Recognising this limitation early is what separates a defensible qualitative or exploratory study from one that overclaims generalisability it cannot support.

Where a usable frame exists, criminology draws on the same probability designs used across the social sciences: simple random sampling, which gives every listed unit an equal chance of selection; stratified sampling, which samples separately within subgroups such as offence category or facility type so small groups are not lost by chance; and cluster sampling, which selects whole groups, such as police stations or courts, and studies everyone eligible within them.

These designs let a researcher calculate a margin of error and generalise back to the frame with a stated level of confidence, which is why prison censuses, court-file audits and victimisation surveys built on an electoral roll or a custody register lean on them.

Where no frame exists at all, which is the ordinary condition for studying active, uncaught offenders, researchers fall back on non-probability designs instead: purposive sampling, which selects respondents deliberately for a trait the study needs; quota sampling, which fills target subgroup counts through whatever channel is available; and convenience sampling, which recruits whoever is easiest to reach.

None of these can support a statistical estimate of a population rate, and each carries its own access-driven bias, but they still assume the researcher can locate and approach eligible people directly.

The full mechanics of these six designs, their variance formulas, their assumptions and the sampling-frame problem in database construction, belong to statistics rather than to criminological fieldwork craft, and are covered in sampling strategies and representative data. What criminology adds, and what the rest of this topic is about, is what happens when even purposive and convenience recruitment fail because the population actively conceals itself: a problem chain-referral sampling and gatekeeper access negotiation were built to solve.

Snowball sampling recruits an initial small group of respondents, often called seeds, who fit the study's criteria, then asks each of them to refer others they know who also fit those criteria.

The sample grows outward through personal networks in successive waves, which is why it works precisely where conventional recruitment fails: it relies on trust that already exists between people in a network, rather than on a stranger's cold approach, to bring in respondents who would never answer a public advertisement or a survey knock on the door.

The technique was formalised as a statistical method by Leo Goodman in a 1961 paper, and its specific value for studying stigmatised and criminal populations was worked out in detail by Patrick Biernacki and Dan Waldorf in their widely cited 1981 methodological paper on chain-referral sampling of heroin users, which set out the practical problems of finding seeds, verifying eligibility and controlling how referral chains spread.

Seed selection matters more than researchers new to the method expect. Seeds drawn from a single clique, a single neighbourhood gang or a single treatment programme will produce chains that stay largely inside that clique, so most snowball designs deliberately start from two or three seeds recruited through unrelated routes, for example one through a needle-exchange programme and one through a probation contact, to give the referral chain more than one starting point to grow from.

Researchers also cap how many referrals each respondent can offer, typically three to five, because a single highly connected seed left uncapped can dominate the entire sample and make it look larger and more diverse than the underlying network actually is.

A further practical problem is verifying eligibility without a frame to check against: a referred contact who claims to deal drugs or belong to a gang cannot be checked against any register, so researchers rely on screening questions, corroboration from the referring respondent, and sometimes a waiting period before a first interview to filter out people who volunteer for a payment rather than because they genuinely fit the study's criteria.

Respondent-driven sampling, developed by the sociologist Douglas Heckathorn in 1997, keeps the chain-referral mechanic but adds structure that plain snowball sampling lacks. Each respondent is given a limited number of numbered referral coupons, the researcher records who recruited whom and how large each respondent's personal network is, and this recruitment and network data is then used to mathematically weight the final sample, adjusting for the fact that better-connected people are more likely to be recruited, in an attempt to produce estimates that approximate a probability sample of the hidden population.

The weighting itself rests on two assumptions that a researcher has to state and defend rather than take for granted: that recruitment approximates a Markov chain, meaning who a respondent recruits depends only on that respondent's own network position and not on the earlier waves that led to them, and that respondents can report their personal network size reasonably accurately.

Both assumptions can fail. Network-size reports are often rounded or estimated rather than counted, and a population with strong homophily, where people overwhelmingly refer others just like themselves, produces recruitment chains that stay inside one segment of the network no matter how many waves run, which the weighting formula cannot detect from the recruitment data alone.

Because it can produce estimates that look like population figures, respondent-driven sampling is often used where a rough prevalence estimate is genuinely useful to policy, for example estimating the size of a city's population of people who inject drugs for a harm-reduction service, or the number of active sex workers in a district for outreach planning.

Criminological studies of offender networks more often use the plain chain-referral mechanic without the coupon and weighting apparatus, because the research question is about mechanism and decision-making rather than a population count, and the added coupon-tracking overhead is not worth the cost when no rate estimate is being claimed.

Neither method eliminates bias entirely. Referral chains tend to stay within existing social clusters, so isolated or peripheral members of a hidden population are structurally less likely to ever be reached, no matter how many waves the chain runs. Respondent-driven sampling's statistical weighting corrects for known network size, but it cannot correct for a whole sub-network the chain never touched in the first place, which is a limitation researchers must acknowledge rather than assume away.

A hidden or hard-to-reach population combines two features: there is no way to enumerate its full membership, and its members have a concrete incentive, legal risk, stigma, or both, to avoid being identified. Active offenders such as burglars and drug dealers, people who use illegal drugs, sex workers, and members of gangs or organised criminal networks all fit this description, which is why criminology has produced most of the methodological literature on chain-referral sampling and fieldwork access.

Reaching these populations usually runs through institutional or informal gatekeepers who control access, a prison superintendent who must approve a research protocol before any interview happens inside a facility, a police commander who decides whether officers may speak to a researcher, a court registrar who controls access to case files, or, in street-level fieldwork, a locally respected figure whose introduction vouches for the researcher's trustworthiness to a wary community.

Negotiating this access is rarely a single approval; it is an ongoing relationship. In the United Kingdom, for example, a researcher who wants to interview prisoners must obtain both institutional ethical approval and separate operational permission from the prison authority before any fieldwork can begin, and that permission can still be withdrawn if the research is seen as embarrassing an institution.

Formal permission from an administrator also does not guarantee genuine cooperation from the people the researcher actually needs to talk to, prison officers or incarcerated people may treat an administration-approved outsider with suspicion precisely because the approval came from above them.

Fieldworkers who study offenders directly, rather than through institutions, instead often need an informal gatekeeper from within the community itself, sometimes a former offender acting as a fixer, whose personal vouching substitutes for institutional permission and can open doors that no letterhead ever could. Ethnographers working in Brazil's urban favelas, for instance, have long relied on this kind of informal community gatekeeper for safe access to neighbourhoods where drug trafficking factions hold effective local control.

Ethics review adds a further, formal layer of gatekeeping in most jurisdictions. Institutional ethics or research boards require informed consent procedures adapted to populations who may be unable to read a standard consent form, who may fear that participation itself is incriminating, or who are legally in custody and therefore under institutional control, a status that research ethics guidance treats as inherently coercive unless special safeguards, such as assuring respondents that refusal carries no consequence for their case or custody status, are built into the design.

Gatekeepers vary in how much discretion they hold. An institutional gatekeeper, a prison superintendent or a court registrar, usually grants access against a written protocol and a paper trail, which makes the terms of access explicit but also makes withdrawal easy to document and justify if the research becomes politically awkward.

An informal community gatekeeper holds no such paper authority, so the terms of access are whatever the relationship implies, which gives the researcher less procedural protection but often faster and deeper access than any institutional route could produce, precisely because the gatekeeper's own standing inside the community is what is actually being lent to the researcher.

Fieldwork with active offenders places researchers in settings that carry real physical and legal risk: unsupervised meetings with people who commit crime for a living, locations chosen by the respondent rather than the researcher, and the possibility of witnessing an ongoing offence.

Standard safety practice includes working in pairs or keeping a colleague informed of meeting times and locations, setting clear rules about where interviews take place, briefing and debriefing after each fieldwork session, and having a pre-agreed exit plan if a situation becomes unsafe.

A separate legal risk concerns what a researcher does with information about ongoing or planned criminal activity disclosed during fieldwork. Most ethics guidance requires researchers to state clearly, before consent is given, the limits of confidentiality, for example that a credible disclosure of planned serious harm to a named victim may have to be reported, so that respondents are not misled about how far confidentiality actually extends.

On generalisability, the honest conclusion is a modest one. Chain-referral and purposive samples describe the people who were reached, who agreed to talk, and who fitted within the network the researcher's seeds happened to connect to. They are not statistically representative of every active burglar or every person who uses a given drug, and even respondent-driven sampling's weighting only corrects for measurable network effects, not for whichever sub-groups the referral chains never entered at all.

Selection bias of this kind means findings from a hard-to-reach-population study travel best as an account of mechanisms, how offending decisions get made, how a criminal network functions, rather than as a precise estimate of a rate or proportion in the wider population.

This does not make the method weak; it makes it a different tool suited to a different question. Where a probability sample of prison records can estimate how many people were convicted of burglary in a given year, only fieldwork with active, uncaught offenders can explain how a burglary is actually planned, how a target is chosen, and how offenders think about risk in the moment, questions a records-based sample cannot answer at all because it never reaches anyone who was not caught.

Access itself is a further source of bias worth naming on its own. A gatekeeper grants access on their own terms, which means the sample a researcher ends up with reflects who the gatekeeper was willing to introduce, not who exists in the wider hidden population.

A prison administration that only approves interviews with prisoners already engaged in a rehabilitation programme, or a community fixer who only vouches for offenders they personally trust, is quietly filtering the sample before chain referral even begins, which compounds the network-clustering bias already built into snowball sampling itself and further narrows what the resulting study can claim to represent.