pseudosamples

Pseudosamples are artificial datasets that imitate the properties of a population or observed data but are not actual new observations. They are created through resampling or generation based on a model or the original data, with the goal of enabling statistical inference, model assessment, or uncertainty quantification without requiring additional real-world data.

Common forms include bootstrap samples, created by resampling with replacement from an observed dataset; jackknife samples, constructed by systematically leaving out one or more observations; and cross-validation partitions, which split data into training and validation pseudo-samples. In time-series or dependent data, methods such as block bootstrap or moving-block bootstrap are used to preserve dependence structure. In predictive modeling, synthetic, or pseudo-samples, may be generated by fitting a model to data and drawing new observations from the fitted distribution, or by techniques such as SMOTE or generative models to augment imbalanced datasets.

Pseudosamples are used to estimate sampling distributions, construct confidence intervals, assess estimator bias, tune models, and evaluate predictive performance. They provide a practical means of uncertainty quantification when true sampling from the population is impractical or expensive.

Limitations include reliance on assumptions about the data-generating process, potential propagation of biases from the original data, and the risk of overfitting or optimistic evaluation if pseudo-samples are not properly independent or if leakage occurs. They are not substitutes for real, external validation data in many contexts.

History: The bootstrap, a principal source of pseudosamples, was introduced by Bradley Efron in 1979, building on earlier resampling ideas such as the jackknife developed by Quenouille and popularized by Tukey.