Adaptive Multistage Sampling
Also known as: AMS, adaptive multi-phase sampling, sequential multistage sampling, adaptive hierarchical sampling
Adaptive multistage sampling combines the hierarchical efficiency of multistage designs with adaptive decision rules that adjust which units are sampled at later stages based on what is observed at earlier stages. It is used when a target characteristic is rare, clustered, or spatially heterogeneous and a fixed design would waste resources on uninformative areas of the population.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use adaptive multistage sampling when the target variable is rare or spatially clustered and a uniform design would yield mostly zero or near-zero observations; when the population is geographically or structurally hierarchical, making a multistage frame natural; and when field data collection is expensive and resources must be concentrated where the phenomenon is present. Do NOT use it when the population is approximately uniformly distributed — the adaptive overhead adds cost with no benefit. Also avoid it when a pre-specified adaptive criterion cannot be defined before fieldwork, or when analysis teams lack the statistical expertise to apply adaptive estimators correctly, as naive use of standard estimators on adaptively collected data will produce biased results.
Strengths & limitations
- Substantially increases the probability of detecting and measuring rare, clustered, or patchy phenomena compared to fixed designs of the same total cost.
- Maintains unbiased probability-based estimation when correct adaptive estimators (Hansen-Hurwitz or Horvitz-Thompson with adaptive inclusion probabilities) are applied.
- Efficiently allocates field effort by concentrating resources on informative primary units identified in real time.
- Flexible: the adaptive criterion and the degree of intensification at each stage can be tuned to the available budget and the rarity of the phenomenon.
- Retains the structural advantages of multistage designs — exploiting natural population hierarchies and reducing travel costs.
- Requires a well-defined adaptive criterion specified before fieldwork; a poorly chosen threshold can eliminate the efficiency gain or introduce bias.
- Final sample size is random and cannot be fixed in advance, complicating budgeting and logistics.
- Estimation is more complex than for fixed designs; standard survey software does not support adaptive estimators directly.
- The design assumes the clustering or rarity pattern is stable during data collection — temporal variation can distort the adaptive allocation.
Frequently asked
Is adaptive multistage sampling the same as adaptive cluster sampling?
No. Adaptive cluster sampling (Thompson 1990) adds neighbouring units around any unit satisfying the criterion within a single stage, forming irregular networks. Adaptive multistage sampling operates within a pre-defined hierarchical frame: it intensifies sampling of sub-units within primary units that meet the criterion, but does not expand to geographically adjacent units outside those primary units. The two designs can be combined, but they are conceptually distinct.
How do I choose the adaptive criterion threshold?
The threshold should be set based on pilot data, prior surveys, or substantive knowledge about what constitutes a meaningfully elevated value of the target variable. A threshold too high means almost no primary units trigger intensification and the design behaves like a standard multistage design; too low, and nearly all primary units are intensified, eliminating efficiency gains. Simulation studies using plausible population models are useful for threshold selection before fieldwork.
What estimator should I use for the population total or mean?
The unbiased estimator for adaptive designs is typically the Horvitz-Thompson estimator, which requires knowing the inclusion probability for each unit under the adaptive rule. Thompson (1992) provides the mathematical derivation. In practice, the Rao-Blackwell improvement of the Hansen-Hurwitz estimator is preferred for its lower variance. Standard weighted expansion estimators from fixed-design software should not be used without modification.
Can I pre-determine the total sample size?
Not precisely. Because the number of secondary-unit observations depends on which primary units satisfy the adaptive criterion — which is data-dependent — the total sample size is a random variable. Researchers typically specify an expected sample size and a maximum budget-driven cap. Simulation under assumed population models can estimate the distribution of final sample sizes to aid logistical planning.
Does adaptive multistage sampling require special software?
Yes, in most cases. Standard survey software packages assume fixed probability designs and do not compute adaptive inclusion probabilities automatically. Researchers typically need custom code — often in R — implementing the Horvitz-Thompson or Rao-Blackwell estimator with adaptively computed inclusion probabilities. The R package ACS provides some relevant functions, though full adaptive multistage support usually requires bespoke scripting.
Sources
- Thompson, S. K. (1992). Sampling. Wiley. ISBN: 978-0471548850
- Cochran, W. G. (1977). Sampling Techniques (3rd ed.). Wiley. ISBN: 978-0471162407
How to cite this page
ScholarGate. (2026, June 3). Adaptive Multistage Sampling. ScholarGate. https://scholargate.app/en/survey-methodology/adaptive-multistage-sampling
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Adaptive Cluster SamplingSurvey Methodology↔ compare
- Adaptive Stratified SamplingSurvey Methodology↔ compare
- Cluster SamplingSurvey Methodology↔ compare
- Multistage SamplingSurvey Methodology↔ compare
- Systematic SamplingSurvey Methodology↔ compare