Why are Counties Difficult to Predict?

A couple of weeks ago Chris Bowers of myDD posted his map of political power by county:


This map is an attempt to show (using the typical red-Republican and blue-Democratic colors) how the country now looks in terms of approval ratings. Using it, Bowers can reasonably claim that the country has gone quite "blue" and he cites the latest SurveyUSA poll to show that in only 3 states does the President have favorable ratings (Idaho, Utah, Wyoming, see map).

Looks encouraging for liberals, right? In fact this map points up some interesting issues to do with politics at levels more detailed than the state. First, the map above is speculation, or perhaps at best extrapolation. As most people know, there are over 3,100 counties in the USA and it's difficult to say how they would look. As comments to Bowers' own post pointed out, his method is unclear (and not likely to be accurate).

Why is this? And why are counties so difficult to predict, politically?

1. There is a lack of data collected at that geographic scale

2. Inferring county-level data from available state data runs into the well-known ecological fallacy.

3. Even if every county was polled, there are many sparsely populated counties in America, which renders the result susceptible to sampling error. We'd also need to know the applicable "sampling frame" or the subset of the population the sample was drawn from (eg., registered voters, random telephone calls, etc.) which also shapes the result.

Perhaps we should forget about this level of data then? No: counties (and geographies at even finer levels of detail) are important:

1. Although many races occur statewide (notably Presidential and Senators) the House of Representatives and of course races for state office are based on districts within the state.

2. Getting an idea of support at the detailed level is important for voter registration and GOTV. You'll want to target areas perhaps even as detailed as the local precinct in GOTV or voter registration.

While I don't know the method used in producing the map above, it wouldn't be impossible to produce a map that estimated political support at the county level, if you could combine good local data/knowledge, previous county-level data (eg., percentage voting Kerry-Bush in '04) and some stratified sampling. I'm thinking of the Census Bureau ACS here.

Important resources:

1. David Leip's Electoral Atlas (purchase data at county level)

2. Anthony Robinson has free voting data he created himself

2. Fairdata2000