|
A pointless exercise? Not necessarily...if you do it reasonably. In one recent thread, Chris Bowers said that he tried to do it in '04 but gave up when huge swings from day to day were required to account for the most minor shifts in the tracking poll numbers.
A recent blog post that got a lot of attention was this one, during the Dem convention, which showed some rather startling results in attempting to divine the daily Gallup numbers: McCain up 6 on Monday, Obama up 8 on Tuesday, followed by Obama up 16 on Wednesday. Talk about a bounce! It fit the reported Gallup numbers, but it sure looked screwy.
What was even more crazy was that if you accepted those particular guesses at the daily numbers, the next day's Gallup tracking number went up by 2 (from 48-42 Obama to 49-41), but to make that work with the prior numbers, Obama's 16 point lead the day before had to vanish COMPLETELY, all the way down into a tie! So what was the problem?
In a word: rounding. The tracking polls report 3-day averages that are rounded to the nearest integer, which means that what looks like a 2 point swing (say, from 45-45 to 46-44) could actually be as little as a 0.4 point swing (from 45.4-44.6 to 45.6-44.4). And when you're guessing daily numbers before averaging them into a 3-day tracking number, if you require integer (i.e. non-rounded) values for each single day, you're magnifying potential rounding errors by several times as much again.
So what I tried to do was go into the Gallup tracker for the last two weeks and get some numbers that made sense, by guessing at daily numbers down to the tenth of a point, i.e. adding one decimal place. So no more guessing that one day's numbers were either 45-45 or 46-44; they could be 45.3-44.6, 45.8-44.2, etc. Then I averaged those numbers and rounded to get the 3-day tracking result.
The results are here, and I'll explain them a bit now. The integer values in the right-most columns are the actual Gallup tracking poll average, with the dates they were reported in the column to the right. The left-most columns are my guess at what the daily numbers have been. The 3-day average for 8/21 came from the daily numbers from 8/18-8/20, the 3-day average on 9/2 comes from the 8/30-9/1 daily numbers, etc.
Now, those daily numbers are guesses, but they're almost guaranteed to be closer to the real numbers than any purely integral values would be. You can fiddle with the daily numbers and still get the same tracking numbers to come out. Each daily number is, of course, restricted by the fact that it contributes to three different tracking poll results, so the permissible "range" for any individual daily number is going to be relatively small, usually no more than 0.2 or 0.3 in any direction. Now, if you change one of those numbers, you'll affect the permissible range for some of the other daily numbers both before and after it, so there's no way to guess the "right" daily numbers.
So does this actually tell us anything? Well...not really. But it might be a good sanity check for anyone who obsesses over these tracking polls too much (like myself, obviously)! |
