Nonprobability Sampling

Nonprobability sampling includes several versions of survey sampling that often are expedient to implement but do not allow calculation of the probability of selection of the sample from among possible samples from a population. The goal of a survey is to gather data in order to describe the characteristics of a population. A population consists of units, or elements, such as individuals, households, businesses, tracts of land, or inventory records. A survey collects information on a sample, or subset, of the population. In some surveys, specific units or elements are chosen by the survey designers to be in the sample. Interviewers are assigned to interview the members of or gather data on the units in the selected sample. Sometimes multiple attempts at contacting and collecting data from the selected sample members are made. In many situations, it is possible to compute the probability that any member of the population is in a sample of a certain size selected with a specified design or protocol. It also is possible in many situations to compute the probability overall that the sample is selected from the population. Such a design is a probability sampling design. Examples of probability sampling include simple random sampling, stratified random sampling, and cluster sampling.

In nonprobability sampling designs, it is not possible to compute the probabilities of selection for the samples overall or, usually, for individuals. Examples of nonprobability sampling include convenience sampling and quota sampling. In convenience sampling, interviewers themselves are given some latitude in selecting the population members to interview or the units on which to record data. That is, the survey designers and planners do not strictly control the selection of the sample. Convenience sampling occurs in many forms, including selection by an interviewer of people at a shopping mall, selection by a waiter of customers at a restaurant, and Internet and call-in polls. Quota sampling is similar to convenience sampling but requires interviewers to collect certain data on certain numbers of individuals from each of several population subgroups or strata. For example, a quota sample could require interviewers to interview a specified number of females and males within age groups in administrative regions within a state. One could implement such a study by randomly calling phone numbers from a telephone book, asking to speak to a person in the household who belongs to one of the available strata, and conducting an interview if possible. If no one is available at a household in an allowable stratum, then the interviewer simply calls the next telephone number.

Estimates of population characteristics based on nonprobability samples are affected by selection bias. Since the interviewers choose respondents that they want to interview, there is a potential for selection bias. If the respondents in the survey are systematically different from the general population on the variables being measured, then estimates of characteristics will be different on average from what they would have been with a controlled probability-sampling scheme. In probability sampling, the survey planner or researcher controls which units are in the sample and selects the sample using known[Page 690] probabilities of selection. The probabilities of selection can be used to produce estimates of population characteristics without the problem of selection bias.

...