Re: [MLUG - DISCUSSION] [POLITICS] another cool graph from Krugman

["Jonathan King" <EMAIL:PROTECTED>]

On Thu, May 8, 2008 at 2:25 PM, Mike Miller <EMAIL:PROTECTED> wrote:
> This does make it look bad for the Republicans:
>
>  http://krugman.blogs.nytimes.com/2008/05/07/phase-two/

How do you figure? This plots results for the past 14 elections.
Incumbent parties have won 7, and the opposition has won 7.

Winning an election is a categorical result. This skanky Pearson's r
stuff predicting vote share misses the point. The real question is
simply this: if you have only a set of binary data point (here, 14 of
them), what decision rule can I create that will best discriminate
winners from losers.

As I mentioned, incumbency provides exactly zero information.

I cannot find any simple cut-point for election-year income growth
that gives me significant predictive power for the outcome of the
election.

I can come up with a rule of the following form: If growth is greater
than about 4%, then the incumbent party wins, or if growth is less
than 2% then the challenger wins; otherwise, it's a coin flip. This is
not a very satisfying predictive rule.

Another rule is: the incumbent party wins unless there is an unpopular
war or an impeachment. That predicts 11 of 14 elections (misses only
1992, 1960, and 1980). If you rephrase that to "military/diplomatic
debacle or impeachment" then you predict 12 of 14 elections (missing
only 1992 and 1960).

Of course, I am data-mining here (as is everybody else, but I
digress). In any case, the last rule suggests there is a better than
80% chance that the GOP will lose in November. But I don't believe it
for a second. There really just isn't enough data here to build any
kind of useful model.

jking





>  Mike
>
>  _______________________________________________
>  discussion mailing list
>  EMAIL:PROTECTED
>  http://mlug.missouri.edu/mailman/listinfo/discussion
>

_______________________________________________
discussion mailing list
EMAIL:PROTECTED
http://mlug.missouri.edu/mailman/listinfo/discussion