Model-Assisted Estimation in Survey Sampling

Model-assisted estimation unites the strengths of design-based survey sampling with predictive models to improve the precision and reliability of population estimates. At its core, this approach employs auxiliary information—variables that are known for all units in the population—to augment traditional estimators such as the Horvitz–Thompson estimator. By fitting a superpopulation model (for example, a linear regression or tree-based predictor) to sampled data and then adjusting design-based estimates with model predictions, survey practitioners achieve lower mean‐squared errors and more robust variance estimates. This synergy is particularly valuable in contexts of unequal probability sampling, small-area estimation or high-dimensional auxiliary data, where naive estimators may suffer from bias or inefficiency. In recent years, developments have spanned de-biasing techniques for machine-learning models under complex designs, unified variance estimation in high-dimensional settings and extensions to functional and spatial data. Globally, model-assisted methods underpin large-scale health and economic surveys, environmental monitoring and official statistics, enabling policymakers to draw more accurate inferences from finite samples while retaining valid design-based inference.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent advances in model-assisted estimation have emphasised integration of machine-learning algorithms with survey design. One study analyses the bias in regression trees and random forests induced by complex sampling, proposing weight‐adjusted splitting criteria and variance estimators that correct for unequal probabilities. The corrected trees not only yield unbiased predictions but also improve interpretability in stratified surveys. Another work introduces a double-objective classification and regression tree (CART) framework for stratum-specific total estimation, balancing classification accuracy for domain membership with regression accuracy for outcome prediction. Applied to health‐outcome domains, ensemble variants demonstrate superior efficiency compared with conventional post-stratification estimators. A third line of research addresses functional data collected in multistage designs, formulating a scalar-on-function regression approach that uses weighted score equations for estimation and novel survey-weighted bootstrap methods for inference. This framework accommodates high‐resolution sensor or time‐series data within national health surveys, yielding valid confidence bands and efficient point estimates under complex designs.

Model-Assisted Estimation in Survey Sampling publication trend

The graph below shows the total number of articles in model-assisted estimation in survey sampling across all publications each year (not limited to Nature Index journals).

Technical terms

Model-assisted estimator: A survey estimator that combines a predictive model fitted to sampled data with design-based weights to improve precision while preserving design-based validity.

Auxiliary variables: Characteristics known for all population units and used in models to reduce variance and correct bias in survey estimators.

Sampling weights: Inverse probabilities of selection assigned to sampled units to ensure design-based unbiasedness and to adjust model fitting in complex samples.

Classification and regression trees (CART): A non-parametric model that partitions data into homogenous subsets for either classification or regression, adapted here to account for survey weights.

Scalar-on-function regression: A modelling framework that relates a scalar outcome to a functional predictor (for example, a time series) collected under a complex survey design, employing weighted estimation and resampling inference.

References

1. Learning de-biased regression trees and forests from complex samples. Machine Learning (2024).
2. Stratum-specific health outcome estimation in Pakistan using double goal CART. PLOS ONE (2024).
3. Scalar‐on‐function regression: Estimation and inference under complex survey designs. Statistics in Medicine (2024).