Here is a bit of a lesson on why you REALLY have to be careful with poll numbers. To understand what a poll is doing, and how well it is done, you have to look at the "internals". The details of the sample, and whether there might be a bias in that sample. Let's take a recent poll:

Newsweek Poll: Republican Convention 2004
Saturday September 4, 11:46 am ET
Bush/Cheney Lead Kerry/Edwards 54 to 43 Percent

http://biz.yahoo.com/prnews/040904/nysa058_1.html

It is of REGISTERED voters, so we don't have to worry about the issues of likely voter models.

Seems to match the TIME numbers, with an 11 point lead. But is that an ACCURATE measure?

If you scroll WAY down you get to the internals:

374 Republicans (plus or minus 6)
303 Democrats (plus or minus 6)
300 Independents (plus or minus 6)

Anyone see a reason why this poll might give a republican ticket an edge? Does this match the reality of political party affiliation?

From http://www.usatoday.com/news/opinion/columnist/neuharth/2004-01-22-neuharth_x.htm

An estimated 201.5 million U.S. citizens age 18 or over will be eligible to vote Nov. 2, although many are not now registered.

Of these, about 55 million are registered Republicans. About 72 million registered Democrats.

About 42 million are registered as independents, under some other minor party or with a "No Party" designation

So, the breakdown of party is 33% rep, 43% dem, 25% ind. (rounded)

The poll has 38% rep, 31% dem, 31% ind. It way undersamples dems, and oversamples reps and inds.

It also has a sample with around 40% military households. Does that sound right to you? Do 40% of households in the USA have someone in the military? I doubt it.

Sooooooooooo, crappy sampling that does not match the population. The sample is biased towards republican respondents. Unreliable numbers.

I am not saying the poll is wrong, but I am saying there is good reason to doubt that it is right.

Then why put such a poll out? If the pollsters know that the sampling is played to the right why do they headline it? I am not of the tin hat variety often, but it seems that the adage of being able to get statistics to prove anything applies in this instance. If you are going to ask more than one republican for every democrat, how can the numbers turn out any other way.

With that then you have to infer that the pollsters seem to have had an agenda?

And it can be summed up as:

Sexy poll results sell.

They get an outlier poll. Was it designed to be biased? We can't know. Let's assume it is just one of those things where the respondants clumped together in a way that did not match the population.

They get the result. Go to press, or go back into the field and sample more people to get the demographics closer to the mean?

One means higher sales, mags or network ratings or whatever.

The other means more money spent, while others are grabbing all the poll press.

Some do go back into the field. Gallup I think it was, after the DNC, got an odd sample. They decided to wait and get more people to respond before they released. (I recall some cries of liberal media bias at the time, though I don't think on the PO board).

Not all news is accurate, and not all polls are created equally. It ticks me off, as a social scientist, but I doubt anything can be done by consumers of information. The only hope is that reporting institutions have some idea what they are doing, and some restraint when the methodology breaks down.