When polling is more like guessing

Election analysts and forecasters depend on accurate polling. Unfortunately, there’s not much of it so far this cycle.

Many of the surveys to date have been conducted by firms that use automated phone surveys and combine deficient sampling with baffling weighting practices.

Nowhere is this more evident than in recent polls in Georgia, a state that should be easy to poll because the secretary of state publishes demographic data on the age and race of registered and actual voters in elections back to 1996…

In contrast with most polls conducted by news media organizations, which collect random samples of adults and then weight to census demographic targets, most automated firms call from lists of registered voters and therefore do not collect a random sample of adults. As a result, they do not weight to census targets for adults, and instead weight to estimates for the likely composition of the November electorate.

That approach can be effective if the pollster can accurately model the electorate, as the Obama campaign did in 2012. But low-cost automated polling firms often operate with only a few staff members, who seem to have less capability to accurately estimate the composition of the electorate.