SAMPLING INFORMATION FOR CONTINUOUS MONITORING

                SAMPLING INFORMATION FOR CONTINUOUS MONITORING

          The telephone survey design for continuous monitoring involved
     RDD sample design issues of some complexity. The sample objective was
     a uniform number of interviews in each "week" (a 17 day interviewing
     period) of a 46 week study period. We should be able to treat each of
     the weekly samples more or less independently. That is, the user
     should be able to take the weekly samples and aggregate them in
     various ways to represent time periods (months, quarters, pre-primary
     intervals) of interest to their research.

          The constraints on the sample design were also clear: time and
     money. Given the strong relationship between time and cost in
     telephone interviewing operations, the chosen sample design had to be
     one which minimized the amount of time that interviewers must spend
     in reaching respondent households. A proven method for reducing
     contact time is the two-stage RDD design originally suggested by
     Warren Mitofsky and Joe Waksberg.

          Very briefly, this design utilizes the A.T.& T. listing of
     telephone central office codes. Each record on the listing is an area
     code-central office code combination, e.g., 313-764; thus each record
     represents 10,000 distinct telephone numbers (e.g., 313-764-0000
     through 313-764-9999). Another way to put it is that each record
     represents 100 clusters of 100 consecutive telephone numbers
     (0000-0099,0100-0199, etc.).

          Each primary selection is one such group of 100 consecutive
     telephone numbers, designated by randomly generating a single 4-digit
     number. For example if (313) 764-4424 is generated, the cluster
     containing numbers 4400-4499 is tentatively designated for selection.
     These primary numbers are called; if they are not working household
     numbers, the clusters in which they fall are not selected. If they
     are working household numbers, their clusters are selected into the
     sample and a specified number of additional four-digit numbers within
     the same cluster is generated. For example, if the desired sample
     size is nine, eight more 4-digit numbers within the hundred series
     would be selected.

          While the Waksberg-Mitofsky method is cost-effective, it
     sacrifices something in precision because of its clustered nature.
     The NES implementation of this design for Continuous Monitoring
     spreads the use of each primary stage sample one-hundred series over
     the 46 week course of the study -- maximizing the distribution of the
     sample and minimizing the clustering effects for short time interval
     analyses. At the end of the 46-week study period, the complete sample
     will contain roughly 700 primary stage numbers (clusters) of 5
     interviews each.

          In the Waksberg-Mitofsky two-stage selection, the several
     numbers selected from each cluster at the second stage are used
     within the same sample period. In the NES variation, each cluster
     that is selected produces one telephone number per week. (This
     telephone number translates into a label of a sample coversheet).
     When there is an interview or some other kind of final disposition of
     the coversheet, the cluster is not used further in the sample week.
     Clusters of primary numbers are in the sample for two weeks, then
     rotate out for 8 weeks. The assignment of clusters produces a 50%
     overlap from week to week. The intent of this overlap is to introduce
     some correlation among observations for short, adjacent intervals of
     time. If successful, the time 1 to time 2 correlations will yield
     improved precision for estimates of change between the two periods.

          As the study design is implemented, it is important to note that
     certain coversheet dispositions mean that a cluster can be
     "re-dialed" within the sample week. For example, if the telephone
     number on the coversheet is of a business (non-household) then the
     next number in the primary number series can replace it.  Other
     redialing situations are non-working numbers or non-sample residences
     (institutions). Some sample coversheets, even though they do not
     yield an interview, cannot be replaced. These include refusals,
     non-interviews of valid respondents, and households with no eligible
     respondents.

          Once a household is reached, selection of respondents within the
     household proceeds by listing all persons within that household
     (male, oldest to youngest; female, oldest to youngest); determining
     which of the residents are eligible (18 on election day, U. S.
     citizens) and using a Kish selection table to randomly determine the
     respondents.

          The method results in slightly unequal probabilities of
     selection. That is, respondents in households with TWO telephone
     numbers are twice as likely to be selected as those with only one
     telephone number. Also, respondents from households with TWO eligible
     residents are only half as likely to be selected as respondents who
     are the only eligible adult.

          The NES staff has compared weighted and unweighted frequency
     distributions for a number of variables. Results indicate that the
     data may be treated as an equal probability sample, i.e., that
     selection weights are not required. Those interested in a further
     discussion of the point should see working paper #5. (See Appendix C).