Sunday, January 19, 2014

Choosing sides not Sides

Booman wants me to set up an account to comment at his tribune. Like hell. He debates John Sides about 2016. I want to pile on.

Associate Professor of Political Science at George Washington University, John Sides, attempts to rebut Dan Balz's fine analysis yesterday that the Republican Party has an uphill battle to win the presidency in 2016. I find his argument unconvincing.

He begins by providing his credentials.

"In April 2012, two other political scientists — Seth Hill and Lynn Vavreck — and I did a presidential election forecasting model for The Washington Post. The model had only three factors: The change in gross domestic product in the first two quarters of the election year, the president’s approval rating as of June of that year and whether the incumbent was running. That model forecast that Obama would win in 2012, and — although there is nothing magic about this model — it was ultimately accurate within a percentage point"
Here not Sides uses only the popular vote.

But later he decides to look at who won the election

Since the passage of the 22nd amendment limiting the president to two terms, only one time (1980-88) has the incumbent party held the White House for more than two consecutive terms. The regularity with which control of the White House changes hands also suggests that the playing field may tip in the GOP’s favor in 2016.
This is Bad statistical analysis. The result of taking a small sample, arbitrarily making it smaller and then deciding which of two variables to average.

The result of the gross data mining is to reject the null that the party that has held the White House has a 50-50 chance at a p level of 7/64 (one tailed) or 14/64 two tailed. I count 6 cases 1960, 1968, 1976, 1988, 2000 and 2008 with one exception to the "rule". This is the sort of data analysis one finds from political scientists and baseball color commentators.

As an aside, Sides's claim of fact is ambiguous and can be interpreted so it is false. The correct claim would be "Since the *ratification* of the 22nd amendment ...." passage can refer to passage by two thirds of the House and the Senate. After the 22nd amendment was passed by congress in 1947, Harry Truman was re-elected in the fifth straight Democratic victory. A statistic which depends on the distinction between congressional approval and passage of an amendment does not pass the laugh test.

But my main point is that, when he isn't grossly cheating, Sides looks at the conditional average popular vote not the outcome of the election. He does this because election outcomes have weird unpredictable stochastic elements (butterfly ballots pregnant chads and such like) which add noise. It is obvious and plain certain that Sides switched from popular vote to outcome to get the (still feeble result) he wanted. Notably, in 2000 the popular vote went the wrong way for him, so the result vanishes even if one scores elections as won or lost based on popular vote (as opposed to averaging popular vote which makes sense). In 1960 the popular vote was an lamost perfect tie. In 1968 and 1976 it was very close. The margins were (with positive supporting Sides) 1960 0.17 1968 0.7 1976 2.06 1988 -7.72 2000 -0.51 2008 7.27 the average was negative until 2008 (see how shameless sample trimming can be) and is now roughly 0.223% (I was hoping for negative based on 1988 which wasn't as horrible as I remember). The standard deviation of the mean is greater than root(22) so it is well over 4. We have a z-score on the order of 0.05. Such almost complete absence of evidence against the null is actually very rare (except when the null is true by definition except for measurement error). Even artificially trimming the sample, the raw data provide no evidence for the 8 year itch hypothesis. It is a pattern found by sifting and resifting a tiny sample. An excellent example of how one should not analyse data.

No comments: