Home Statistical Dictionary Design-Based Inference
Design-Based Inference
An approach to survey inference in which the finite population's values are treated as fixed, unknown constants, and all randomness in an estimator's properties comes solely from the physical sampling design used to select the sample.
In Plain English
Design-based inference is the classical foundation of survey sampling theory: it treats every value in the population as a fixed number that simply happens to be unknown, there's no assumption that anyone's income or test score was 'randomly generated' by some underlying process, the only randomness in the whole framework is which specific units happened to get drawn into the sample. All the standard machinery, unbiasedness, standard errors, confidence intervals, in classical sampling theory is justified purely by that physical act of random selection, which is exactly why a probability sample supports rigorous inference without needing to assume anything at all about the shape or distribution of the outcome variable itself.
Definition
Design-based inference treats the values of the finite population, , as fixed, unknown constants, and derives the statistical properties of an estimator (bias, variance, confidence intervals) entirely from the randomness induced by the sampling design, the specific probability mechanism used to select which units enter the sample, repeated over hypothetical repetitions of that same random selection process. This is the classical framework underlying Horvitz-Thompson estimation and the finite-population-correction formulas taught in introductory sampling theory; its key strength is that it requires no assumption whatsoever about the distribution or functional form of the outcome variable in the population, unbiasedness and valid standard errors follow purely from the known probability sampling mechanism, which is what makes probability sampling designs the gold standard for defensible population inference, but its scope is limited to quantities directly estimable from the design (it offers no natural way to 'borrow strength' across related domains the way model-based inference does), which is precisely the gap that model-based and model-assisted methods like small area estimation are built to fill.
Properties
- Design-based inference's freedom from distributional assumptions is its central appeal, a design-based confidence interval's validity rests entirely on the known, controlled randomization of the sampling design, not on any assumption about how the outcome variable is distributed in the population, which is a substantially weaker and more defensible requirement than most model-based statistical methods need.
- The classical Horvitz-Thompson estimator, and the standard finite-population-correction formulas for simple random sampling, are canonical design-based estimators, their unbiasedness proofs never reference any assumed distribution for y, only the known probabilities built into the sampling design itself.
- Design-based inference has no natural mechanism for improving precision in domains or small areas with very few sampled units, since it deliberately avoids leaning on any model of the outcome, this is exactly the situation model-based inference and small area estimation are designed to address, by explicitly trading the model-free guarantee of design-based inference for extra precision borrowed from a statistical model.
At a Glance