Canadian Forest Service Publications

A second look at endogenous post-stratification. 2015. Magnussen, S; Andersen, H.E.; Mundhenk, P. Forest Science. 61(4):624-634.

Year: 2015

Issued by: Pacific Forestry Centre

Catalog ID: 36205

Language: English

Availability: Not available through the CFS (click for more information).

Available from the Journal's Web site.
DOI: 10.5849/forsci.14-183

† This site may require a fee

Mark record


Properties of the poststratification (PS) estimator of variance are evaluated in simulated sampling and 252 settings (3 populations × 4 levels of stratification × 3 sample sizes × 7 methods of stratification) of endogenous poststratification (EPS). Poststratification was done on the basis of predictions from a linear or a nonlinear model fitted to the sample to be stratified. Empirical variances obtained with regression estimators (REG) were generally significantly smaller than EPS variances. The efficiency of a stratification method depends on the population but only marginally on the number of strata or sample size. Optimization of strata boundaries (three methods) leads to a slightly higher empirical variance and a concomitant and mostly significant underestimation of the expected variance. EPS with regression was not more efficient than REG alone. A practical calibration of the biased estimates of variance does not seem possible. For the purpose of map-updating and map-consistent statistics, a nonoptimized EPS holds practical advantages over the regression approach.

Plain Language Summary

Post-stratification is a widely used in forest inventories to improve overall efficiency and generate sample estimates for specified domains of interest. When post-stratification is based on sample dependent estimates of strata-membership of population elements, the stratification is called endogenous post stratification, or EPS for short. Although some studies have suggested that EPS can be approximately as efficient as a regular post-stratification based on information external to the sample, our study confirms that EPS will give the user overly optimistic estimates of uncertainty if the classifier for predicting post-strata membership has been optimized to achieve the best fit to the sample data. In short, EPS is not as efficient as post-stratification.