Rasmussen Tracking Poll [Obama vs McCain]

[muon2]

Seems a little weird that it's all tied up with the favorables being so different. Interesting.

Obama will likely be up by around a point or two tomorrow, as a good McCain sample comes off tomorrow.

The trick to figuring out whether this is a "bounce" or actual movement, once again, is probably going to be by figuring out whether any major shift in party ID has occurred that Rasmussen is weighting away.  

My best advice for figuring out whether this lasts beyond the next week or so is simply to compare Gallup to Rasmussen.  If Gallup continues to poll 2-4 points ahead for McCain after about a week or so, then this may be some type of actual movement in the party ID.  If they move back to tracking each other, like they did pre-DNC, then it isn't.

The difference between favorability ratings and voting preference may well be an indicator of the party ID shift. I don't know the extent if any that party ID is used in the calculating the favorables, but I presume that there is some difference. If the weight is less hard, then I would expect movement in the favorable number better reflects shifts in the raw sample.

I'm surprised that Rasmussen only recalibrates the weighting monthly at this time. I would think that a semimonthly adjustment would make more sense after Labor Day.

[muon2]

Last five days:

              M+3        O+8      O+12      M+1       O+6      O+10   
Obama 50.44% (50.43%/49.81%/49.95%/49.47%/49.40%)
McCain 45.05% (44.35%/44.37%/45.16%/45.96%/46.56%)

Sample today is *roughly* McCain +2.07.  Rowan may come in and correct me later on this number.

EDIT:  My sleepiness makes it not McCain +2.07%, but rather McCain +3.22%.  Ugh.

Thats just really really silly. Of course there is noise within a tracker, but O+10 to M+1, back to O+12 and then down to M+3 is completely illogical. If those are the actual results (and we know ras doesn't release them) then there is something badly badly wrong with his polling. But I don't believe those are his figures and thats just someone somewhere (maybe you?) taking the piss. There will be 1 to 3 point movement from day to day, thats all - backed up by the tracker only moving 1 or 2 points every day.

One test for the daily tracking numbers is to measure their fluctuations compared to the expected fluctuations given the margin of error. For a daily sample of 300, and assuming relatively little change in the real voter preference, 19 out of 20 daily samples should be within 5.7% of the mean. The six daily numbers shown above have an average of O +5.3, and three results are in excess of 5.7 from that average. Based on that I would be suspicious.

On other thing:  

If there's a problem in the reverse engineering of daily samples from three day averages, it would most likely manifest itself in an error in the daily estimates having period three (meaning that an error in one day of a three day average would automatically propagate itself every three days down the line).  

To put it differently:  Random noise will tend to not have any recognizable cycles.  Methodological error will have a cycle of length three.  If your data has a conspicuous cycle of that length, there likely is an error somewhere. 

This is the natural place to be suspicious, and though there are only six days, there is a clear cycle. I would agree that the statistical spread and the visible cycle lead me to think that the initial data going into the reverse engineering of the average were off. That error is now propagating forward as large daily fluctuations.

[muon2]

The only thing I would point out, muon2, is that I separately calculated the averages of the one-day samples using the internals of the separate one-day only polls used in:

1) Debate results (Saturday Prez sample)
2) Bailout numbers (Sunday Prez sample)

When Rasmussen gives these numbers, he also gives the internals of how they break down in GE w/leaners (McCain v. Obama).  

While not entirely accurate (he does not include the small number of undecideds), we can use these internals multiplied to the actual support of the other numbers to make an educated guess at the end number.  So, this adds a bit of a cross-check into the analysis.

For Saturday, I said Obama +7 to 8 (the formula has it at 7.70%)
For Sunday, I said McCain +2 (the formula has it at 3.22%).  Looking at the numbers again now, and doing the calculations, McCain +3 actually fits more accurately with the internals.

I'm not saying your estimate is wrong, I'm just saying that the variation is suspiciously large. I recognize I only have 6 points to deal with, so this set can be a statistical fluke. However, I think the occurrence of three points outside the 95% range, when I expect less than one gives me reason to raise a question.

[muon2]

The only thing I would point out, muon2, is that I separately calculated the averages of the one-day samples using the internals of the separate one-day only polls used in:

1) Debate results (Saturday Prez sample)
2) Bailout numbers (Sunday Prez sample)

When Rasmussen gives these numbers, he also gives the internals of how they break down in GE w/leaners (McCain v. Obama).  

While not entirely accurate (he does not include the small number of undecideds), we can use these internals multiplied to the actual support of the other numbers to make an educated guess at the end number.  So, this adds a bit of a cross-check into the analysis.

For Saturday, I said Obama +7 to 8 (the formula has it at 7.70%)
For Sunday, I said McCain +2 (the formula has it at 3.22%).  Looking at the numbers again now, and doing the calculations, McCain +3 actually fits more accurately with the internals.

I'm not saying your estimate is wrong, I'm just saying that the variation is suspiciously large. I recognize I only have 6 points to deal with, so this set can be a statistical fluke. However, I think the occurrence of three points outside the 95% range, when I expect less than one gives me reason to raise a question.

Nah, I understand your point.  My main concern has to do with whether the beginning of the sequence is incorrect or not.  If it is, then the entire sequence would be wrong.  All we could do is tell, vis-a-vis the three-day average, that McCain's numbers improved 2.09% (roughly) in comparison to yesterday's numbers.  We couldn't tell the *starting off point*, which would allow us to determine how this movement compares to the sample that dropped off (which would then be able to give us the actual daily sample).

This is one of the reasons why I'm starting to do the internal one-day checks on Rasmussen's other questions that he asks that are released separately in the polling database.  I want to see how close my guesstimates come to the ladder sequence.  So, for the last two days, these have been my guesstimates (based on the internals):

Sat (debate sample):  O53-M45
Sun (bailout sample):  M50-O47

They might be a tad lower on both sides (more undecideds), but considering the fact that Friday's sample was most certainly a high undecided sample (both candidates moved downwards in raw internals, clearly), I have to believe that this is a reasonable guess.

The best test would be an extended time correlation analysis. If data over the last 3 weeks shows a strong three day cycle then the initial data is likely off. The good part of this type of analysis is that it can pick up cycles independent of an underlying long-term trend.

I'm not saying your estimate is wrong, I'm just saying that the variation is suspiciously large. I recognize I only have 6 points to deal with, so this set can be a statistical fluke. However, I think the occurrence of three points outside the 95% range, when I expect less than one gives me reason to raise a question.

but this is statistics, clustering is always a possibility.

Clustering is a statistical possibility, but a persistent three-day cycle would indicate a systematic bias in the process. This assumes that short term effects should be statistical, and there is no external environmental effect to provide voter shifts in a three-day pattern.