But suppose that we wanted to sample from all persons that had used the library in the past year. In most cases there is no such list of users, especially if users may be people who are not registered borrowers and persons from outside our normal service area. In a school setting where almost every student will have been a library user we might settle on sampling from the student body and then eliminating the non-users. In a public library setting there is no dependable list of people in the community. Even if there were we might have to contact a large number of people to find just a few users. In fact we may conclud that there is no reasonable way to sample previous users no matter how earnestly we may desire to do so.
Another case in which sampling may not work is where the the identification codes are not in a nice sequential order like our numbered list of employees above. For example, imagine the trouble of trying to generate call numbers at random. For Dewey numbers what would we do? We might generate three digits before the decimal and three after for the class, a letter and three digits for the Cutter number and three letters (including blanks so as to generate one and two letter combinations) to identify the title. Such a design would simply not work. The vast majority of the generated "call numbers" not only would not be in our collection, they would represent numbers that had never been assigned in the entire history of Dewy Cataloging!
Some populations cannot be sampled by random sampling simply because there is no way of going from a randomly chosen designator to specific members of the population. Other populations cannot be sampled because we have no way of finding their members. While sampling is a powerful tool, it is not always appropriate or even feasible. Ultimately we must remember that the reason for using random sampling was to give us a means of justifying the accuracy of our sample result. Many research techniques in psychology, sociology, and anthropology use alternate methods of verifying that results obtained from a small population can be appropriately discussed in the context of a large population.
this page is at http://testbed.cis.drexel.edu/sample/notto.html