- A group can be hard to reach for four different reasons: it is rare, it has no list, it is hard to locate, or it refuses; each reason calls for a different solution.
- When a group is rare, a general survey is expensive in screening; disproportionate stratification where the group is concentrated cuts the cost without giving up probability sampling.
- Multiple lists and time-location sampling allow estimates with known probabilities, provided the corresponding weights are calculated.
- Snowball sampling finds people but does not produce a representative sample; respondent-driven sampling supports estimates only under conditions that must be checked.
- Language, interview mode, the interviewer and trust decide who answers as much as the sample design does.
1. The problem: the groups that matter most are missing from the data
A programme to reintegrate returning migrants, a service for older people living alone in villages, an information campaign for a linguistic minority, or a product for Moldovans working abroad all share one problem: the people they need to understand are exactly the people an ordinary survey fails to capture. A national sample of 1,000 contains too few members of the group for separate analysis, and the ones it does contain are often the easiest to find, and therefore the least typical. Hard-to-reach populations is the umbrella term for these groups, and researching them calls for a different design, not just a bigger sample.
This article explains why a group becomes hard to reach, which sampling designs exist, what each can generalise, and how fieldwork is organised. It is a practical companion to the guide on public opinion surveys, which covers the ordinary case of a general population with a good sampling frame.
2. Hard-to-reach populations: four different problems
The phrase covers four situations that are often confused, although they need different solutions.
- The group is rare. It is a small share of the population, for example people with a particular disability or households with a member who recently left to work abroad. They can be found, but a great many households must be checked to identify them.
- The group has no list. There is no sampling frame: no register, no customer base, no membership list. This is the typical situation of the diaspora, informal workers and small unregistered farmers.
- The group is hard to locate physically. People are mobile, work shifts, live in remote localities or are away part of the year. Older people in small villages, seasonal workers and commuters belong here.
- The group is hard to persuade. Members avoid strangers, fear consequences or have had bad experiences with institutions. Here the problem is not the sample but trust.
Most real groups combine two or three of these problems. The first design step is to establish which one dominates, because a good solution for a rare group (more screening) does nothing for a group that refuses.
3. When a general survey is enough, and what screening costs
If the group has a known and not very small share of the population, the most robust solution is still a probability sample of the general population, with screening questions at the start of the interview. The cost arithmetic is simple. A group that makes up, hypothetically, 5% of households requires on average 20 contacts for every eligible interview; 300 interviews need about 6,000 households contacted, before nonresponse. If the group is 1%, the cost multiplies fivefold.
Screening also has its own error: respondents must acknowledge, at the door or on the phone, that they belong to the group. Eligibility questions must be simple, neutral and free of any hint of benefit; otherwise some eligible people declare themselves ineligible to avoid the interview, and others do the opposite. The eligibility rate achieved is compared with any external source available.
4. Probability designs for rare groups
Three designs reduce the cost of screening without giving up known selection probabilities.
Disproportionate stratification
Rare groups are seldom spread evenly. If administrative data or a census show that the group is concentrated in certain districts or localities, these become strata with a higher sampling fraction. The gain is real only when the differences in concentration are large, several-fold; with small differences the advantage disappears. The results are then combined with design weights, a procedure described in the article on weighting survey data.
Multiple lists
Sometimes partial lists exist: beneficiaries of a programme, members of an association, users of a service. None covers the whole group, but together they cover a large part of it. The sample is drawn from each list, and a person who appears on two lists has twice the chance of selection. The correction is either to remove duplicates across the lists before selection, or to apply a weight equal to the inverse of the number of lists on which the person appears. Removing only the duplicates that happen to turn up in the sample does not correct the bias. The part of the group that appears on no list remains uncovered and must be estimated separately or declared.
Time-location sampling
For groups with no list but which frequent known places, a frame of locations and time slots is built: a market on Fridays, a bus station at weekends, a community centre in the evening. A random sample of location-time combinations is drawn, and in each one eligible people are interviewed according to a fixed rule. Estimates are valid for the population that frequents these places, not for those who avoid them, and frequent visitors have a higher chance of selection, which is corrected by asking how often they visit.
5. Network sampling: snowball and respondent-driven sampling
When there are neither lists nor venues, the group's social networks are used. In snowball sampling, a few initial respondents are asked to refer other members of the group, who in turn refer others. The method finds people who would otherwise be very hard to find and is valuable for qualitative and exploratory research. It does not, however, produce a representative sample: those referred resemble those who refer them, well-connected people are over-represented, and isolated people are almost entirely missing.
Respondent-driven sampling is a controlled version of the same idea. Each respondent receives a small number of coupons, usually two or three, to pass on to other members of the group; the study records who recruited whom and how many group members each respondent knows. Recruitment continues over several waves, and estimates are weighted in inverse proportion to the size of each person's network. Respondent-driven sampling was developed for stigmatised or hidden groups and can produce estimates, but only when several conditions hold: the group forms a connected network, ties are reciprocal, respondents report their network size accurately, and recruitment has run for enough waves that the sample composition no longer depends on the starting respondents. These conditions are checked in the data and reported; where they do not hold, results are treated as indicative.
6. Fieldwork: language, interview mode and trust
With hard-to-reach populations, the organisation of fieldwork matters as much as the sample design. A few practical rules:
- Language. The questionnaire is administered in the respondent's language, with translated and verified versions, and interviewers speak it fluently. A multilingual team is not a logistical detail but a condition of data quality with linguistic minorities.
- Interview mode. Older people in rural areas are almost always interviewed face to face, through CAPI, because many households have no internet and some have no working phone. The diaspora, by contrast, is hard to reach face to face and is more often studied online, with all the representativeness limits of an online survey.
- The interviewer. For groups wary of institutions, an interviewer from the community or introduced by a local leader achieves much better response rates. On the other hand, they may influence answers to sensitive questions, a risk handled through training and supervision.
- Timing. Seasonal workers are at home in certain months, migrants around the holidays, shift workers at unusual hours. The fieldwork calendar adapts to the group, not the other way round.
- Trust and data protection. The purpose of the study, confidentiality and the right to refuse are explained clearly, and identifying data are kept separate from answers. For vulnerable groups, the rules are set out in the guide on research ethics and data protection.
7. Analysis: what can and cannot be generalised
Figures for hard-to-reach populations are read according to the design that produced them. A probability sample with screening or disproportionate stratification supports estimates with a margin of error, after weighting and with the design effect calculated. Multiple lists and time-location sampling support estimates for the covered population, provided selection probabilities are reconstructed. Respondent-driven sampling supports estimates only if its assumptions are checked, and its confidence intervals are usually much wider than those of a random sample of the same size. Snowball sampling and recruitment through adverts support descriptions and internal comparisons, but not percentages that generalise.
A useful rule: if the decision needs a figure for the whole group (for example, how many potential users a service has), the design must be probability-based. If the decision needs an understanding of needs and barriers, a mix of qualitative interviews and a network sample is often enough and far cheaper. For groups where differences between women and men are essential, the design must secure adequate separate bases, as the guide on gender-sensitive research explains.
8. Comparing the designs
| Design | When it is used | What it can generalise | Main limit |
|---|---|---|---|
| General survey with screening | Rare but not very rare group, no list | The whole group, with a margin of error | High contact cost; misreporting of eligibility |
| Disproportionate stratification | Geographically concentrated group | The whole group, after weighting | Little gain unless concentration is strong |
| Multiple lists | Overlapping partial lists exist | The population covered by the lists | People on no list; duplicates must be corrected |
| Time-location sampling | Group frequents known places | The population that frequents those places | Excludes those who avoid them; frequency correction |
| Respondent-driven sampling | Hidden group linked by ties | The whole group, only if assumptions hold | Assumptions hard to verify; wide intervals |
| Snowball | Exploration, qualitative research | Does not generalise percentages | Sample dominated by well-connected people |
With hard-to-reach populations, the choice is not between "right" and "wrong" but between degrees of certainty and cost. What matters is that the report states clearly which rung the results stand on.
9. What to ask the agency
- The operational definition of the group and the eligibility questions, tested before fieldwork.
- Why the group is hard to reach (rare, no list, mobile, distrustful) and the proposed solution for each reason.
- The sample design, with the covered population named explicitly and an estimate of the part left uncovered.
- How weights are calculated: for strata, multiple lists, visit frequency or network size.
- For respondent-driven sampling: the number of coupons, the number of waves, the checks on assumptions and the confidence intervals.
- The questionnaire languages, the translation procedure and the interviewer profile.
- Consent, confidentiality and data protection procedures compliant with the GDPR and Moldova's Law no. 133/2011.
- A limitations section in the report stating which figures can be generalised and which are only indicative.
Hard to reach does not mean impossible to research. It means the design, fieldwork and analysis are built around the group, and that claims of representativeness match the design. If you are preparing a study of migrants, minorities or isolated communities, our social and political surveys team can help you choose the design that fits the decision and the budget.
See also: Social and political surveys Multilingual team CAPI and CATI Weighting survey data Face-to-face interviews (CAPI) Gender-sensitive and inclusive research
Do you have a decision to take?
Tell us the problem, not the method. We propose the right research design and explain what it can and cannot deliver.