Gallup's Explanation of their Tracking Poll Procedures
From a 10/12/00 email by David W. Moore of Gallup to AAPORNET
(listserve of the American Association for Public Opinion Research).
Used with permission.

Several comments have been made about the CNN/USA Today Gallup tracking poll, and I thought it might be helpful to clarify some of the procedures that we use. 

The tracking poll is based on what is essentially a continuous sampling design, with some new sample added each night and some old sample called back each night.  Because of the continuous nature of the design, we are in fact able to make numerous call-backs.  While we have set up a minimum 5-call design, in fact for many respondents we make more than the 5 calls required by the design before they are dropped. 

During the past five weeks (through last Sunday), Gallup completed interviews with 14,228 respondents. The percentages of interviews completed on each call are shown below:

31.4% of interviews completed on 1st call
21.4% on 2nd call
14.5% on 3rd call
10.7% on 4th call  
7.1% on 5th call  
5.1% on 6th call  
4.1% on 7th call  
5.7% on 8th call or higher

The CASRO response rate for the continuous sampling since it began in September is currently 40%.

The Gallup likely voter model assumes that 50% of our national sample of adults will turn out to vote.  Each respondent who is registered to vote (or says he/she does not need to register) is administered a set of likely voter questions that results in a 7-point likely voter scale, where each respondent gets a score of anywhere from zero (a default assigned to those who are not registered, or who do not intend to vote) to 7 (most likely to vote). Based on that scale, we include 50% of the respondents into the likely voter model.  Typically, that includes all of the respondents who score a 7, and a proportion of those who score a 6. 

Thus, for example, about 40% of the national sample scores a 7 and 17% scores a 6 (hypothetical -- but not far off reality on some days).  That means we need to get 10% of the national sample out of category 6 to go with the 40% in category 7 to make up the total of 50%. Since we actually have 17% of the sample in category 6, we weight that category with .5882 (10/17) and we weight category 7 with a '1' -- while categories 0 through 5 get a weight of zero.  The net result is a weighted sample that represents half of the national sample.

When we calculate the number of respondents included as "likely voters" we count the full number of respondents included in categories 6 and 7, even though we typically weight the category 6 respondents to about .5 or so. That means that our raw number of respondents on whom we base the likely voter results is typically about 700 or so over a three-day  period.

Interest in the election and likelihood of voting are two measures that help constitute the likely voter scale, so when interest wanes among one party and surges in the other (as can happen during a party convention, or in the wake of a debate where one candidate is seen as much better), the proportion of Democrats and Republicans in the sample of likely voters can vary as well.  Gallup does not weight its sample by party -- sometimes Republicans outnumber Democrats by several percentage points and sometimes Democrats outnumber Republicans by several percentage points.  If we weighted by party, the fluctuations we show in the ballot would be considerably dampened, but we do not know for certain what the actual distribution will be on election day and are reluctant to assume it will be the same as in other elections.  If on election day, one party or the other has been particularly energized or, conversely, de-energized, our model would reflect that fact.  If it turns out that competition is so intense that both parties are equally energized, that should be reflected in our model as well.  

David W. Moore
The Gallup Organization
47 Hulfish Street
Princeton, NJ 08542