I haven’t followed the Washington trial over the vote in the Governor’s race very closely, so I was surprised to read that ecological inference was even being attempted given the circumstances.
The problem of ecological inference is one of the bigger problems in trying to determine how people vote in political science. Simply put, given individual vote choice isn’t available to scientists other than as exit polling, one has to utilize precint level voting returns as the lowest level of analysis to study voting behavior (different though correlated to surveys of political beliefs).
The problem is that one cannot reliably infer the behavior of individuals from a group behavior. If all you have is information concerning the aggregated group, one cannot make inferences about the individuals within the group or how they behave.
To give a fairly simple example, if one has an integrated precinct with 50% black and 50% white voters, one is unable to make any meaningful analysis of vote choice by race unless everyone votes in one way. If the final vote choice is heterogenous, one doesn’t know how individiuals voted and that means one couldn’t determine how people vote by race in such a case without an exit poll.
The reason we know a lot about black voting behavior is that America is, in a scientific sense, conveniently segregated so precincts with large numbers of African-Americans tend to only have African-Americans (please–no examples of your neighborhood–my precinct is about 50-50 and it’s an abberation).
So, the Rossi camp finds Jonathan Katz, a very talented methodologist at the California Institute of Technology to argue for an idea called proportional reduction in dealing with votes by felons that shouldn’t have occurred. He suggested simply taking the number of votes in such precincts and then reducing the number proportionally for all of the likely illegal votes in that precinct for both sides. So if Gregoire got 60% of the vote, take the total number of likely illegal votes and reduce her total by 60% of the illegal votes and reduce Rossi’s vote by 40% of the total likely illegal votes.
It’s attractive to Republicans because it reduces her totals more than his totals.
The problem is that it is statistical incompetence and so the judge ruled this morning that commits the ecological fallacy that one can determine individual behavior by aggregate statistics.
Katz’s argument was essentially that since there is limited information, making a choice based on that limited information is the best one could do–so being ‘fair’ was by assuming that felon votes were entirely independent of any factor other than geography. Such an assumption is ludicrous–it contains no information about race, gender, income, church attendance or a wide array of demographics that affect vote choice.
The argument that any information–such as geography in this case–should be used to determine how to count votes in such a case is one that essentially leads to a basic point in how we count votes–namely not counting votes.
The simple answer is that with the information available, there is no way to fairly distribute the votes between candidates. Without individual information or a large amount of specific information about the precincts and people in them where actual ecological inference techniques could be used with at least some ability to gauge the reliability of the stats, no court and no political scientist should be making conclusions about a fair distribution of those votes. To do so is statistical incompetence. One might argue for a particular hypothesis, but no one should be able to make the argument that they have reasonably confident conclusions based simply on geography.
Even with more data, one would have to be especially cautious. As Gary King, the political scientist who literaly wrote the book on ecological inference says:
Because the ecological inference problem is caused by the lack of individual-level information, no method of ecological inference, including that introduced in this book, will produce precisely accurate results in every instance.
What Katz would do is make an assumption about voters who were felons and how they voted with no information except geography to inform that assumption. For Republicans to argue for such a methodology is bizarre given their complaints over well understood techniques for oversampling in relation to the Census which would be on far better statistical ground than Katz’s idea.
While such techniques may have been utilized in other cases, that doesn’t make the technique anymore reliable or reasonable, it makes such cases as warning signs against junk science.