VA/NJ governor races not predictive

Underscoring a point I've made several times in the last week, Alan Abramowitz, a respected political scientist at Emory University, has a new analysis showing that the Virginia and New Jersey gubernatorial results are not predictive of midterm election seat changes:

[T]he results of the previous year's gubernatorial elections in Virginia and New Jersey did not predict the results of the midterm elections. Not only is the estimated coefficient for the Virginia/New Jersey election variable small and statistically insignificant, but it is in the wrong direction: the better Republicans did in Virginia and New Jersey, the worse they did in the subsequent midterm election.

A forthcoming paper in Legislative Studies Quarterly (via John Sides at The Monkey Cage) finds that changes in partisan control of House seats in special elections is predictive of general election outcomes. In the current context, these results suggest an environment that is slightly more favorable to Democrats due to the pickup in NY-23 (the only partisan change in a House special election this electoral cycle).

In reality, of course, Democrats will have a tough time in 2010 -- I fully expect them to lose a significant number of seats. But contrary to the media hype, the Virginia and New Jersey results don't provide much information about what that outcome is likely to be.

Comments

[MartyB]

Brendan –

After reviewing Abramowitz columns, I think you state your conclusions with too much certainty.

As I understand it, Abramowitz main point is that the party of the President is the main contributing factor of mid-term results. His regression analysis shows this and anyone would have to agree.

However, even he points out that :

“On occasion, victories in Virginia and New Jersey have been followed by big gains in the midterm elections. In 1993 Republicans won both of these off-year contests; one year later they gained 54 House seats and 8 Senate seats and took control of Congress for the first time in almost half a century. Twelve years later, in 2005, Democrats swept to victory in Virginia and New Jersey; a year later they picked up 30 House seats and 6 Senate seats and took control of Congress back from the GOP.”

So it’s possible under certain circumstances, there may be a correlation.

So the issue is not whether victories in VA and NJ *always* tend to indicate results in mid-terms (his analysis shows they don’t), but whether on the tails (i.e. in certain perhaps extraordinary circumstances) they have predictive power.

I don’t see any indication in his analysis that he controlled for the first effect (president’s party) before measuring the predictive power of the governor’s races, although since he didn’t provide much detail, he may have.

The danger of ignoring the distribution tails was well illustrated in last year’s financial meltdown which was at least partially caused by investment managers assuming that their regression-based models were *always* right just because they were right inside the risk distribution tails.

Skepticism is indeed in order (not something the popular press is very good at, esp. if it confirms biases) but your conclusion of "no predictive power" appears far from certain to me.

[Jinchi]

So it’s possible under certain circumstances, there may be a correlation.

Yes, but the point Brendan is making is that the races aren't predictive.

Your example ("On occasion, victories in Virginia and New Jersey have been followed by big gains in the midterm elections...") is exactly what we would expect if we were judging the predictive value of a coin toss. On occasion, big gains would follow a heads-up, and on occasion they wouldn't.

If that never happened, we'd consider the Va/NJ result to be predictive of Democratic gains in 2010. If it always happened, we'd predict Republican gains.

But we can't do either. So it's not predictive.

[Rob]

Jinchi, "always" and "never" set too high a standard to judge whether something is predictive. The magnitude of the coefficient of correlation and its statistical significance reveal how reliably predictive something may be; certainty is not required for a measure to be usefully predictive.

[Jinchi]

Jinchi, "always" and "never" set too high a standard to judge whether something is predictive.

That's true. But as the paragraph Brendan quotes points out the correlation coefficient is small and statistically insignificant (0.09 over 11 election cycles). You could easily get the same correlation coefficient by tossing a coin.

And it still wouldn't be predictive.

[Miles]

Everybody is overthinking midterm performance.

Let's say a given district has 65% turnout for the presidential election, with a slight in-district victory for the Dem candidate, who wins the election. The Dem representative also wins, because Dems won the district.

Things happen, some good for Dems, some bad. Two years pass.

Now it's the midterm. Turnout is 45%. Of the 20% dropoff, most of it are people who were excited to vote for the president, but not for the rep. That means that the remainder is mainly Republican, which means the rep loses.

There's almost nothing to be done about this, short of Obama having a personal referendum. Why? Because people are, quite naturally, more excited about the president than a representative--especially if th rep has done nothing exciting personally.

[MartyB]

Jinchi - Higher correlation does mean more predictive power, in general.

One of my questions was whether Abramowitz controlled for the variable with the most predictive power (party of President) before calculating the correlation of the NJ and VA Governor’s races. My statistics knowledge is a little rusty, but I seem to recall that not all regression techniques are appropriate for analyzing the power of secondary variables. In other words, it's possible the low correlation calculated for the governor’s races is because the technique was flawed.

My other point was that regression analysis is reliable only to a certain degree and only around a somewhat normal distribution/circumstances. In other words, the governor’s races in those years may be meaningful predictors in "non-normal" years (which if they happen as much as even 1 out of 11 instances might indicate the situation in these states is indeed not normal).

As my last comment above made clear, skepticism is required in both directions - i.e. “they are predictive” vs. “they are not predictive”.

[Jinchi]

skepticism is required in both directions - i.e. “they are predictive” vs. “they are not predictive”.

See, this is where we disagree. Here's the prediction

The election of Party X in NJ and Va Governor's races mean that Party X will likely pick up seats in Congress next election

From the 7 elections listed in which both states elected a member of the same party that prediction was correct 3 times, and wrong 4 times. Focusing on the NJ race only: It was correct 5 times out of 11, wrong 6 times out of 11. Focusing on the Va race only: It was correct 6 times and wrong 5 times.

These are exactly what you'd expect from flipping a coin.

Brendan and Abramowitz point out that there is a prediction that is highly correlated.

Members of the president's party lose seats in the midterms

That statement was true 9 times out of 11 the same elections.