I wrote on this topic a few years ago, but here is an updated (and hopefully nonpartisan) consumer guide to reading political opinion polls.

The overall theory of polling is sound. It is entirely possible to predict, within certain relatively small boundaries (and subject to the possibility of an “outlier” sample), the overall tendency of a large group by sampling a smaller, representative sample. Thus, if there were a million rocks in a pile of different colors, picking a sample of just 1000 of the rocks, if the sample is representative, would normally lead to sound conclusions about the relative number of rocks of each colors in the entire group. However, it is crucially important that the sample represent a random sample of the entire group.

However, sometimes polls are not random. The classic example of a political poll that did not produce a representative sample was the 1936 Literary Digest poll in the presidential election between President Franklin D. Roosevelt and Alf Landon, his Republican challenger. The Digest sent out survey cards to 10 million people, using lists of telephone and car registrations. About 2 million were returned and indicated an overwhelming victory for Mr. Landon.

Unfortunately for the Digest, which folded a few years after the poll, it was President Roosevelt who won in a landslide. There were actually two errors in the polling technique. First, the response was self-selected (people getting the cards could choose to return them or not) and second, the cards were sent to a disproportionate number of people with telephones and cars. In the Depression, this was a group that tended to vote Republican. Needless to say, reputable modern pollsters do not use the Digest’s techniques.

In the context of election surveys, it is important to get right the partisan makeup of the sample. This is quite important because partisans do vote for their party in very high numbers; in most elections, upwards of 80 percent of both Democrats and Republicans vote for their party’s candidates. Some surveys directly control for this variable (i.e., directly adjust the sample to get the right mix), while others do not. While the partisan breakdown is fairly steady over time, making appropriate adjustments can be extremely important. For example, in 2008, pollsters had to take into account the large numbers of young people participating in the election, usually supporting Barack Obama. Be very suspicious of a survey that has too many Republicans or Democrats in the sample.

Also of importance is whether an election survey is of adults, registered voters or likely voters. At least 40 percent of voting age adults do not vote. Registered voter election polls are better and likely voter surveys even better, assuming the screening technique of likely voters is reasonable. Generally, likely voter polls produce somewhat more favorable results for Republicans.

Of course, bias is a very important factor in evaluating a poll. Wording of questions can be crucially important is how people answer. In particular, it is well to be suspicious of internal polls commissioned by a candidate, which well use sampling techniques designed to produce favorable results.

Finally, keep in mind that polls do drift up and down statistically, without meaning anything. A tracking poll that shows President Obama’s approval at 49 percent one day and 47 percent the next may not be showing any significant pattern. In addition, although pollsters downplay this, even with good statistical technique, one poll in 20 will be outside the statistical margin of error.

Again, political polls are a valuable tool. But caveat emptor.