List-Assisted Sampling

Impetus for a New Method

In the United States, only 20% of all possible telephone numbers are assigned to a residence. This produces problems for researchers conducting surveys with the U.S. population via telephone in that the amount of work that would be needed for interviewers to cull through all the phone numbers in the country is enormous. A telephone sample with only 20% of the numbers reaching the targeted household is extremely inefficient, and as a result, the survey costs increase as does the length of the field period. There have been various sampling methods used in the past to address this problem, but by far the best known is [Page 430]the Mitofsky-Waksberg method. Introduced in the late 1970s, the Mitofsky-Waksberg method of RDD takes advantage of the fact that residential numbers are often clustered together consecutively in the BCR telephone database, which contains information on telephone exchanges and their geographical mapping throughout the United States. By drawing a sample of 100-banks, which consists of the area code, the prefix, and first two digits of the four-digit suffix, and then dialing a randomly selected number within these banks to determine whether residences are contained within the bank, the Mitofsky-Waksberg method culls the nonresidential sample at a much more efficient rate. A 100-bank will only be retained if the random number that is dialed is indeed a residential number; otherwise, the 100-bank is discarded. Once the 100-banks are chosen, then the telephone numbers are generated by assigning a random two-digit number to the end of the 100-bank exchange.

The two-stage RDD sample design is not without its problems, however. First, some clusters of 100-banks may not contain the minimum number (k) of residential numbers required for that bank. Hence, this greatly slows down the efficiency of calling, as all numbers in this bank must be called in order to meet the minimum number of phone numbers. Second, determining the residential status of a number by simply dialing that number is not necessarily a foolproof method. Often the status of the number will be unknown, and the cluster may be rejected erroneously during the first stage of the sample design. Also, the person responding to the initial phone call may regard the number as a residential number, when in reality it may be something other than a residential number, which would then mistakenly make the 100-bank eligible for inclusion. Third, each cluster must be monitored throughout the field period to ensure that k numbers are sampled from the cluster. This is a great drain on resources and often results in longer field periods as cluster yields may only become apparent later on in the data collection period. Further, numbers used as replacements for nonresidential numbers within a given cluster will not receive as many chances for resolution as those numbers identified as residential numbers early on in the field period. Lastly, the most cumbersome problem with the Mitofsky-Waksberg method is the two-stage cluster design it utilizes, which increases the variance of the estimates when compared to a simple random or stratified design. As these problems made themselves more apparent over the course of experience with telephone surveying in the 1980s, a new method was sought.

...