Fist of all, metanalysis is not my field. I have never engaged in that type of study and never studied it in school nor have I read more than a few in journals. Thus, my opinions on the subject and techniques are pretty suspect.
If I were going to do a metanalysis I would have to make several assumptions. First, that each analysis was drawn from the same metapopulation. This means that differences among them are due to sampling error, not due to structural differences in the population. Then I think I would need some population parameters and also the means and standard errors of the polls. From those parameters I could bootstrap each analysis and resample them as many times as I wished (I think that's how most of the polls work) and from that test for differences between polls. If no difference then combining the means is not a problem. If any are different then we would need to go into the actual data and look for causes of differences.
In other words, the only time we would need to know specifics concerning characteristics of the polled population is in order to describe statistically significant differences between the polls.
Here's what I guess that RCP does. I'll bet they assume that when a polls say something like, "All registered voters" that they assume all of the above criteria are met concerning population and statistical parameters. So they simply take a linear mean.
But like I said, it ain't my field and I don't care enough to find out just how wrong I am. Every time I have tried to fool with resampling statistics I got bitch-slapped.
Usually in meta-analysis, a weighted average is used, not a linear mean.
The classic work in the field:
Meta-Analysis in Medicine and Health Policy
Meta-analysis is one of the primary tools in this little hobby of mine:
Data mining - Wikipedia, the free encyclopedia
When it comes to political polls, they are all tending towards inaccuracy. They predicted Kerry would be president, and this year, they failed in primary after primary as Romney and Guiliani would attest. The meta-analytical problem that needs to be solved is why all are erroring on the same side: poll after poll showed Kerry the winner by a few points. All were within the margin of error, but the skew made the conclusion that Kerry would win seem obvious. The conclusion one reaches is that all the polls are making some unknown invalid statistical measurement, or the election was fixed.
My guess is that the allowance for cell phones is off. I rarely even answer my land line, and I'm thinking of just dropping it. My cell phone is provided by my business, and is not listed to me personally. There must be myriad others who are doing the same thing. Since income is a predictor of cell phone use, my guess is the polls are wrongly skewed towards Obama since his population would be less likely to own cell phones and if they did they would be directly registered to them. I have a gut feeling that the unpolled critical population to the voting sample, upper income cell phone users, would be in a 50/50 split since most information workers are upset about Republican increases of H1-B quotas and have the most to fear from out-sourcing, so I would give that population to Obama, while the remainder, business professional, would probably go to McCain, so this unmeasured skew is probably moot this time around, while in 2004 the general distaste for Kerry among the professional class caused his loss.
The inability to solve this problem for the polling agencies is insurmountable - if most of these people are using phones supplied by their employers or firms they own, it is virtually impossible to reach them given current methods.