Cluster Sampling vs Stratified Sampling

At a Glance

These two approaches solve different problems. Cluster sampling makes data collection affordable when your population is spread across a large area. Stratified sampling ensures you can say something meaningful about specific subgroups. Most large M&E surveys use both.

When Cluster Sampling Saves Money

Cluster sampling groups your population into natural units (villages, schools, health facilities, districts) and then randomly selects a subset of those units. You survey everyone in the selected clusters, or a random subsample within each.

The logic is simple. Sending enumerators to 30 villages and surveying 15 households in each is far cheaper than sending enumerators to 450 randomly selected households scattered across 200 villages. You visit fewer locations. Travel costs drop. Supervision is easier.

Worked example. A nutrition survey covers a region with 500 villages and 50,000 households. Simple random sampling requires visiting households across potentially all 500 villages. Cluster sampling selects 40 villages and surveys 20 households per village (800 total). The team visits 40 locations instead of potentially hundreds. Field costs drop by 60-70%.

The tradeoff is precision. People within the same village tend to be more similar to each other than to people in other villages. They share the same market, the same health facility, the same water source. This similarity, measured by the intra-cluster correlation (ICC), means each additional person in the same cluster adds less new information than a person from a different cluster. You need more total respondents to achieve the same precision as simple random sampling.

The design effect quantifies this cost. If the ICC is 0.05 and each cluster has 20 people, the design effect is 1 + (20-1)(0.05) = 1.95. You need roughly twice the sample size of a simple random sample. At ICC 0.10, the design effect jumps to 2.9. Use the Sampling Calculator to compute this for your specific parameters.

When Stratified Sampling Ensures Representation

Stratified sampling divides your population into mutually exclusive groups (strata) based on a characteristic you care about, then samples independently within each group. Common strata in M&E: urban vs. rural, male vs. female, program districts vs. comparison districts, income quintiles.

The logic is different from clustering. Stratification is not about cost. It is about guaranteeing that your sample represents key subgroups and that you can make valid comparisons between them.

Worked example. A livelihoods program operates in 3 agroecological zones: highland, lowland, and coastal. The program theory predicts different outcomes by zone. Simple random sampling might, by chance, undersample the coastal zone (smallest population). Stratified sampling allocates a minimum sample to each zone, ensuring you can analyze results separately. If you need to detect a 10-percentage-point change in each zone with 80% power, you might need 150 households per zone, for 450 total. Without stratification, you might need 600+ total to ensure enough coastal households appear by chance.

Stratification improves precision when strata are internally homogeneous. If households within a stratum are more similar to each other than to households in other strata, stratified sampling produces smaller confidence intervals than simple random sampling with the same total sample. This is the opposite of clustering, where within-group similarity hurts precision.

Stratified Cluster Sampling: The Common Hybrid

Most large-scale M&E surveys use stratified cluster sampling. This is not an either/or choice. The two approaches combine naturally.

How it works in practice:

1. Stratify the population by the characteristic you need to compare (region, urban/rural, program/comparison).
2. Within each stratum, list all clusters (villages, enumeration areas, schools).
3. Randomly select clusters within each stratum, using probability proportional to size (PPS) so larger clusters have a proportionally higher chance of selection.
4. Within each selected cluster, randomly select a fixed number of respondents (typically 10-25 households).

Worked example. A baseline survey for an education program covers 4 provinces (strata). Each province has 80-120 schools (clusters). The design selects 15 schools per province (60 total) and surveys 20 students per school (1,200 total). Stratification guarantees provincial-level estimates. Clustering makes it feasible to reach 1,200 students without visiting all 400+ schools.

Sample size calculation for the hybrid. Start with the sample size you would need for simple random sampling. Multiply by the design effect to account for clustering. Then allocate across strata. For proportional allocation, each stratum gets a share proportional to its population. For equal allocation (when you need stratum-level estimates with equal precision), each stratum gets the same number. See How to Choose Sample Size for the full calculation.

Design Effect: The Number That Matters

The design effect (DEFF) is the ratio of the variance under your actual sampling design to the variance under simple random sampling. It is the single most important number in cluster sampling. Ignore it and your confidence intervals are wrong, your significance tests are invalid, and your conclusions are unreliable.

Typical design effects in development evaluations:

Rule of thumb: If you do not have a local ICC estimate, use 0.05 for health indicators, 0.10 for education and economic indicators, and 0.15 for attitudinal indicators. Always err on the high side.

Common Mistakes

Mistake 1: Ignoring the design effect entirely. This is the most common sampling error in M&E. A sample of 384 households (the classic "infinite population" calculation) spread across 20 clusters does not give you the precision you think it does. If the DEFF is 2.0, your effective sample size is only 192. Your confidence intervals are wider and your statistical tests are weaker than reported. Always calculate the design effect and adjust your sample size accordingly.

Mistake 2: Too few clusters with too many respondents per cluster. Precision in cluster sampling depends more on the number of clusters than on the number of respondents per cluster. Surveying 50 people in 10 villages gives you less precision than surveying 20 people in 25 villages. Once you pass 20-25 respondents per cluster, adding more people in the same cluster adds very little information. Spend your budget on more clusters, not more interviews per cluster.

Mistake 3: Stratifying on too many variables. Every stratum you add subdivides your sample further. Stratify by region (4 strata) and urban/rural (2 strata) and you have 8 strata. Add gender and you have 16. If your total sample is 800, each stratum gets 50 respondents, which is often too few for meaningful analysis. Stratify only on variables where you genuinely need stratum-level estimates or where the difference between strata is large enough to affect your overall estimate.

Mistake 4: Using cluster sampling when you can afford simple random sampling. If your population is concentrated in a small area (one city, one district), the cost savings of clustering are minimal but the precision loss is real. Use simple random or stratified random sampling when logistics allow it.

Mistake 5: Forgetting to account for non-response. Inflate your sample by 10-20% to account for households that are absent, refuse, or are unreachable. In cluster sampling, losing an entire cluster (security incident, road washout) is devastating because you lose all respondents at once. Plan for 1-2 replacement clusters.

Mistake 6: Treating disaggregation as stratification. Disaggregation means breaking down results by subgroup after data collection. Stratification means designing the sample to ensure adequate subgroup representation before data collection. If you need reliable subgroup estimates, stratify. If you just want to report overall results broken down by group, disaggregation of a well-designed sample may be sufficient, but check that subgroup sample sizes are adequate.

Decision Guide

1. Is your population geographically spread out?

• Yes, and field logistics are a major cost driver: Use cluster sampling (or stratified cluster sampling).
• No, population is concentrated: Use simple random or stratified sampling. Skip clustering.

2. Do you need to compare subgroups?

• Yes, with reliable estimates for each: Stratify by those subgroups. Allocate enough sample to each stratum for the analysis you plan.
• No, you just need an overall estimate: Stratification is optional but may still improve precision.

3. How many subgroups do you need to compare?

• 2-4: Stratified sampling works well. Each stratum gets enough sample for meaningful estimates.
• 5+: Consider which comparisons are most important. You cannot stratify on everything with a finite sample.

4. What is your budget?

• Enough for all locations: Simple random or stratified. Maximum precision.
• Limited (must reduce travel): Cluster sampling. Accept the precision tradeoff and increase sample size to compensate.

Use the Sampling Calculator to compute sample sizes for any combination of these approaches.