One of the main problems encountered by a researcher when designing an Internet survey is the sampling procedure. In this case the classic sampling methods are not effective.
The construction of a good sample is critical to creating a valid survey. According to Kish (1987) and Groves (1989), it is possible to define four different types of populations1:
Population of Inference (PI): The set of individuals that are the objects of the study, in a defined time interval (e.g. the population living in Italy in the first semester of 2011).
Target Population (TP): The finite set of individuals target of the study. It is possible that some units belonging to the PI are not included in the TP (e.g. people in the hospital at the moment of the survey). The difference between the PI and TP is defined by the researcher and is generally driven by practical purposes (e.g. the researcher can choose to not survey people living on high mountains).
Frame Population (FP): A list of units used for drawing the sample. It should be a list of elementary units, for example individuals, or a group of elementary units like families.
Survey Population (SP): The set of individuals that will be, if selected, polled. Obviously, the individual inclination to participate at the survey is unrelated with sampling strategies.
There are differences in the size of each of the populations produced by some biases that can be present at the beginning, at the end or during the survey. Without pretending to be exhaustive, some sampling biases are:
Noncoverage: The list used to identify sampling units is incomplete; this situation generates the “noncoverage error” (the difference between TP and FP).
Total Nonresponse: Some units do not respond to the survey (unit nonresponse) and this situation generates the “unit nonresponse error” (the difference between FP and SP).
It is very important to separate the four populations in order to minimize errors. Obviously, noncoverage error is more serious than unit nonresponse error because the latter is measurable.