Answers are grounded in this paper and cite their sources.
Stratified sampling is considered, where (a) the mean integrated squared error (MISE) metric is used in place of the mean squared error (MSE) metric; (b) the entire distribution [i.e., f(x)], rather than a property of the distribution [e.g., E(x)], is used as a target of the procedure; (c) the distribution f(x) is estimated by a truncated series (to counterbalance model complexity with sample size availability); and finally, (d) samples are taken both with and without replacement. In the last regard, series term inclusion rules are generalized to deal with samples taken without replacement from a finite population. Two special cases are treated in detail. The first shows that even for sample sizes as small as three and for data forms as elementary as the bivariate binary vectors, the use of orthogonal series representation can lead to smaller expected error than would be achievable through use of conventional representation and methods. The second demonstrates that in the continuous case, MISE-based and conventional sample size selection rules can differ substantially.
No takes yet. Share an insight, caveat, or question.
Michael E. Tarter (1986) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: