Posted: Thu Nov 25, 2010 6:42 pm
Ed, Don:
Yes there are problems with both (sampling a predefined area versus 'random' samples'). I do either or both under different circumstances. Often I use a stratified random or stratified systematic sample, which sometimes has intermediate characteristics.
With regard to 'random' samples, in a forest setting random samples do not guarantee independence of the samples, and systematic samples (e.g. set out on a grid) are more accurate for most purposes, and are actually more independent for the same average distance between plots. You have to remember that there is no statistical test that requires a random sample, they require independent samples (even non parametric methods), and we make the assumption that random samples gives independence, and get so stick on random samples that we forget that independence of samples is what we are after. Since forests are patchy, one can measure the level of spatial dependence by distance, and then choose to place plots far enough apart so that the variables of interest one one plot have no influence on the others nearby--then the samples are independent and a systematic sample will be easier to carry out and more accurate.
Then, also, don't forget the tradeoff between statistical significance and ecological significance. With 1000 plots, one can establish statistically significant differences that are tiny and ecologically/biologically trivial. People (even scientists, but not me) do this all the time, and publish meaningless results.
Lee
Yes there are problems with both (sampling a predefined area versus 'random' samples'). I do either or both under different circumstances. Often I use a stratified random or stratified systematic sample, which sometimes has intermediate characteristics.
With regard to 'random' samples, in a forest setting random samples do not guarantee independence of the samples, and systematic samples (e.g. set out on a grid) are more accurate for most purposes, and are actually more independent for the same average distance between plots. You have to remember that there is no statistical test that requires a random sample, they require independent samples (even non parametric methods), and we make the assumption that random samples gives independence, and get so stick on random samples that we forget that independence of samples is what we are after. Since forests are patchy, one can measure the level of spatial dependence by distance, and then choose to place plots far enough apart so that the variables of interest one one plot have no influence on the others nearby--then the samples are independent and a systematic sample will be easier to carry out and more accurate.
Then, also, don't forget the tradeoff between statistical significance and ecological significance. With 1000 plots, one can establish statistically significant differences that are tiny and ecologically/biologically trivial. People (even scientists, but not me) do this all the time, and publish meaningless results.
Lee