Sampling for small NGO surveys: a plain guide
Choose between simple random, systematic and cluster sampling, understand sample sizes used by UNHCR and WFP, and avoid common sampling mistakes.
Short answer
A sample is a selection of households or people that represents a larger group. Use random selection when you have a list of households, and cluster sampling when you do not. Decide your sample size from the precision you need. UNHCR uses 360 households for 5% precision in a camp, and WFP suggests 150 to 250 households for each reporting group.
Key points
- A good sample is selected by chance, not by convenience.
- Random selection needs a list or a way to count households.
- Cluster sampling is used when no list exists; it needs a larger sample.
- Each group you want to report separately needs its own sample.
- Write down how you chose the sample, so you can repeat it at endline.
What is a sample, and why use one?
A sample is a selection of households or individuals from the whole group you want to learn about. WFP's food security assessment guidance explains that it is rarely possible or necessary to survey every household. If the sample is representative, you can generalise from it to the larger population. If it is not, your numbers mislead you.
Small NGO surveys need the same discipline. A baseline of 60 households chosen because they were near the road does not show what the whole target area looks like.
Which sampling method should you use?
UNHCR's household survey guidance shows the options. Choose by what you know about the households.
| Method | How it works | Needs | Best when |
|---|---|---|---|
| Simple random | Every household has an equal chance of selection | A complete, current list of households | A list exists, such as a registration list |
| Systematic random | Pick a random start, then every nth household | A list or accurate count of households | A list or count exists, and an even spread is wanted |
| Cluster | Randomly pick clusters, such as villages, then households inside them | A list of clusters | No household list exists, or villages are far apart |
UNHCR notes that for simple and systematic random sampling you need an up-to-date list of households, or a way to derive or count them. If no information on households exists, cluster sampling will likely be used.
How many households do you need?
There is no single number. Two public rules of thumb:
- UNHCR (camp surveys): about 360 households for random sampling at 5% precision, 100 for 10% precision when resources are very limited (this does not allow comparisons within the camp), and 210 for cluster sampling, using 30 clusters.
- WFP (food security surveys): a sample of 150 to 250 households for each reporting group, ideally nearer the top. For cluster designs, about ten households per cluster is usually enough, so at least 15 clusters for 150 households and 25 clusters for 250.
WFP also notes that the sample size applies to each reporting group. If you want separate results for displaced and host households, or for two districts, you need a sample for each.
Cluster designs need larger samples than simple random designs for the same precision. These figures come from specific agency contexts. Ask a MEAL adviser to confirm the right size for your survey and population.
How do you draw a simple random sample in five steps?
- Define the target group. For example: all households in the 12 project villages.
- Build the list. Use registration lists, or count households. Give each a number.
- Set the sample size. Use your rule of thumb or a calculator.
- Pick random numbers. Use a spreadsheet function such as RANDBETWEEN, or draw numbered papers.
- Record the process. Save the list, the method and the final sample, so you can rebuild it at endline.
What about qualitative work?
Focus groups and interviews do not aim for statistical representation. UNHCR notes that qualitative methods usually use purposive sampling, selecting people by the characteristics or knowledge needed, in order to understand the depth and variation of views. Choose groups by sex, age and other relevant differences. See focus group discussions and key informant interviews.
How do you handle access limits?
In areas where staff cannot visit all villages, be honest about coverage. If you can reach only some places, say that the results describe only those places. Do not present them as results for the whole area. Where phone surveys are safe and suitable, remember that people without phones are left out. Describe these limits in your report, and see remote monitoring.
What mistakes should you avoid?
- Convenience sampling. Choosing nearby or easy households creates bias.
- No list and no plan. Decide the method before you go to the field.
- Too small for the groups you want. Each reporting group needs its own sample.
- Changing the method at endline. Use a comparable method so changes are real. See the baseline guide.
- Not recording non-response. Track refusals and absences.
How EMPO can help
EMPO is a trusted, Yangon-based training and consultancy provider and one of Myanmar's leading names in practical NGO training, with 1,200+ people trained and 15 partner organisations. Our MEAL and intensive support training covers survey design, sampling and data tools, in English and Burmese, online or in person across Myanmar. See our services or contact us.
Sources
- Sampling Methodology and Size — UNHCR WASH Manual
- 2.7 Sampling — World Food Programme, Emergency Food Security Assessment Handbook
- Selecting Participants for Qualitative Data Collection — UNHCR, 2026
Questions people ask
How many households do I need to survey?
It depends on the precision you need and the design. UNHCR commonly uses 360 households for random sampling at 5% precision, and 100 at 10% precision where resources are very limited. WFP suggests 150 to 250 households per reporting group as a rule of thumb.
Can I ask community leaders to choose whom to interview?
For a survey that must represent the population, no. Selection by leaders or by convenience adds bias. For focus groups and interviews, purposive selection is normal, because the aim is to hear different views, not to produce statistics.
What if some selected households are absent?
Return at a different time, then use a pre-agreed replacement rule and record each case. Do not replace households by choosing the easiest ones.