In simulation sampling from forest populations using sample sizes of 20, 40, and 60 plots respectively, confidence intervals based on the bootstrap (accelerated, percentile, and t-distribution based) were calculated and compared with those based on the classical t confidence intervals for mapped populations and subdomains within those populations. A 68.1 ha mapped population, constructed out of 0.081 ha (1/5 acre) Forest Inventory and Analysis of the USDA Forest Service (FIA) plot measurements at 2 points of time in Maine, United States and a 64 ha mapped population from Surinam, South America with only one measurement, were used. The plot designs used were the FIA plot consisting of a 1-ha circular plot subsampled by four 0.0169 ha (1/24 acre) subplots, a 0.081 ha plot that was used in the north east by FIA, and a 10-point cluster of VRP plots that was used in the north central by FIA.
The confidence intervals of all estimates, even those from a sample of size 60, failed to meet the nominal 95% standard. Of the 4 methods used to derive the intervals, the classical method was consistently best in terms of achieved confidence level. This was followed by the three bootstrap methods: t-distribution based, accelerated, and percentile.
We recommend use of the classical t confidence intervals even for small sample sizes and less common attributes. If the classical standard error cannot be computed easily, the t-distribution based bootstrap should be used since it gives only slightly less reliable confidence levels. Also, the bootstrap standard error can often be computed in situations where the classical standard error cannot be.