I wonder how pseudoreplication (https://en.wikipedia.org/wiki/Pseudoreplication) differs from resampling (which is simply resampling with replacement from a given sample).
On the one hand, resampling methods offer a convenient way to assess the distribution of statistics (which is good).
On the other hand, pseudoreplication seems to have a really bad reputation and is stated to be a problem. I understand that pseudoreplication was pointed out as a problem because of repeated measurements of an individual which are clearly correlated, while performing statistical tests etc. which assume uncorrelated data. ("Pseudoreplication occurs when observations are not statistically independent, but treated as if they are" https://bmcneurosci.biomedcentral.com/articles/10.1186/1471-2202-11-5#:~:text=Pseudoreplication%20occurs%20when%20observations%20are,correlated%20in%20time%20or%20space.)
However, I found wordings like
- "pseudoreplication techniques, especially bootstrap ..." https://www.google.de/books/edition/Practical_Methods_for_Design_and_Analysi/reYkGe9PLMQC?hl=de&gbpv=1&dq=%22pseudoreplication%22%20%22bootstrap%22&pg=PA149&printsec=frontcover&bsq=%22pseudoreplication%22%20%22bootstrap%22
- "The next step in the logic is that of pseudoreplication. This involves resampling the data after they are collected but in a way that reflects the sampling procedures" https://support.sas.com/resources/papers/proceedings/proceedings/sugi22/STATS/PAPER279.PDF
which suggest that resampling techniques produce pseudoreplications.
Additionally, I found this book which seems to discuss pseudoreplication due to boostrapping on page 272 (but I do not get the point of the argument)
If pseudoreplication is such a big problem, then why doesn't this imply that resampling always has to suffer from this problem of pseudoreplication? And if this pseudoreplication is a problem in the bootstrap, is there a nice illustration showing the problem of psueodreplication and when do I know that I am running into the problem of pseudoreplication when applying the bootstrap?
I am struggling with "pseudoreplication" being branded as an incredible huge problem whenever you search for it, but also resampling (which is really cool) being called a pseudoreplication method. I want to get the implications right here.
PS: having a set of paired feature vectors and labels, i.e. ((feature_1, feature_2, ...), label) it is clear to me that it only makes sense to do a pair bootstrap and resample the whole tuple (feature vector, label) because each feature_i might be correlated to another feature. But applying the bootstrap here to resample single feature values is an obvious application error.
PPS: This post has more literature on pseudoreplication: Is Hurlbert 1984 the best introductory overview to pseudoreplication?