Network News

X My Profile
View More Activity

Structural losses

By Karl Smith

Kevin Drum suggests that Democrats performed even worse than the 45-seat loss the Hibbs model would have forecast on the basis of structural factors alone.

But the model I wrote about, which comes from Douglas Hibbs, only predicted a 45-seat loss, and it looks like Dems are likely to lose at least 60 seats. That means Democrats underperformed the Hibbs model by 15 seats or so, which is a record for them. ... They've underperformed by ten seats a couple of times in the postwar era, but never by more than that. So at the same time that it's correct to blame most of their losses on structural factors, it's also correct that this was something of a historically bad result.

However, this is the wrong way to interpret those results. The model was estimated using all elections until this one. By design, the computer tries to come up with an equation that gives the least possible error in all previous years. The question is: If we rerun the model using the 2010 data, does 2010 still look like an outlier or will the model now simply predict that the economy was a bigger factor than it thought before?

Also, as a note, Hibbs uses real disposable income growth, which doesn't make as much sense to me as change in unemployment.

Karl Smith is an assistant professor of economics and government at the University of North Carolina and a blogger at ModeledBehavior.com.

By Karl Smith  | November 3, 2010; 5:30 PM ET
 
Save & Share:  Send E-mail   Facebook   Twitter   Digg   Yahoo Buzz   Del.icio.us   StumbleUpon   Technorati   Google Buzz   Previous: The message
Next: Reconciliation -- and more

Comments

>>Hibbs uses real disposable income growth, which doesn't make as much sense to me as change in unemployment.>>

Real disposable income affects everyone, unemployment only has a direct effect on the unemployed.

Neither is perfect, but I'd go for income growth.

Posted by: fuse | November 3, 2010 5:45 PM | Report abuse

More fundamentally, are you suggesting incorporating yesterday's results into the model, then seeing how yesterday compares to the model?

If you average in an outlier, of course the model will move closer to the outlier.

Am I misinterpreting your post?

Posted by: fuse | November 3, 2010 5:49 PM | Report abuse

that's good fun saying that their losses were due to "Structural issues" meaning that it had little to do with the ruling party's policies.

It simply isn't true.

I don't know any Republican that actually blames the economic downturn on Obama. What they blame him for is the butt-headed, jobs-killing legislation that had to do with everything other than jobs, which is what we expected.

No one expected the Dems to spend a whole year on HealthCare, especially when the economy was the real issue. They also didn't expect this administration to spend all that political capital on cap 'n tax and all the other Democrat wet dreams.

The order of the day was jobs, jobs, jobs...and they simply screwed the pooch.

Posted by: WrongfulDeath | November 3, 2010 5:56 PM | Report abuse

What exactly should Obama have done about jobs?

Reducing healthcare costs helps employment and helps reduce the deficit. It's essential for the long-term health of the US economy.

Posted by: fuse | November 3, 2010 5:59 PM | Report abuse

"However, this is the wrong way to interpret those results. The model was estimated using all elections until this one. By design, the computer tries to come up with an equation that gives the least possible error in all previous years. The question is: If we rerun the model using the 2010 data, does 2010 still look like an outlier or will the model now simply predict that the economy was a bigger factor than it thought before?"

Please please please keep this in mind the next time you post on how well ARRA is doing.

"Also, as a note, Hibbs uses real disposable income growth, which doesn't make as much sense to me as change in unemployment."

http://econlog.econlib.org/archives/2010/05/the_econometric.html

Posted by: justin84 | November 3, 2010 6:06 PM | Report abuse

@fuse. Yea, I think he's confused about how regression works. It makes no sense to fit to data and then go back and make predications on that data, it's a recipe for over-fitting and illogical interpretation. You would absolutely move closer to yesterday's result. And that would be meaningless. Perhaps we should include an election where Democrats win by 100% and see if that moves the model in their favor?

Now, you can argue that, since an economy this bad and losses this large have never been found previously, we were extrapolating rather than interpolating (always risky) and our economics-voting models simply weren't effective in this range, and therefore you can't really tell whether Republicans over or underperformed.

But personally I think they overperformed, largely due to voter confusion and anger about bailouts, stimulus, and public spending.

Posted by: CarlosXL | November 3, 2010 6:27 PM | Report abuse

The standard error in the Hibbs model is 11 seats, so his actual prediction was 45 (+/- 11). In this case, we should keep in mind that the actual results were only one standard deviation out.

Regression models have limits: they are not infallible psychic predictions. In this case, we can speculate that the coefficients need refining, as Smith suggests (Hibbs acknowledges that a one standard deviation change down in the incumbency coefficient would result in a Democratic pickup estimate of 199 rather than 222), or that exogenous shocks simply pushed the results further out than estimated (as was likely in the 2002 midterms).

Posted by: TWBr | November 3, 2010 6:29 PM | Report abuse

'Reducing healthcare costs helps employment and helps reduce the deficit. It's essential for the long-term health of the US economy.'

Except nobody is actually seeing reduced health care costs. What they see is just a bunch of discredited liberal politicians talking about future reduced healthcare costs.

Posted by: krazen1211 | November 3, 2010 6:38 PM | Report abuse

Of course if we reestimate the model using 2010 data, it will give a larger estimate for the economy effect in its effort to incorporate the 2010 outcome. The economy is an independent variable, the election outcome is the dependent variable, and what models do is try to match the IVs to the DV. The economy-effect will be larger, and the match with the 2010 outcome will be better.

But that won't answer the basic question: was the pre-2010 model in error due to underestimating the economic effects, or because there was an exogenous shock (eg, Obama screwing up, or the Tea Party)? WIthout lots of data for out-of-sample testing, such a question is basically unanswerable. But thinking that reestimating the model using 2010 data would answer it is a serious mistake!

Posted by: Ulium | November 3, 2010 6:50 PM | Report abuse

>>Except nobody is actually seeing reduced health care costs.>>

That's because right wing opposition prevented things from taking effect for a while.

Posted by: fuse | November 3, 2010 6:59 PM | Report abuse

"That's because right wing opposition prevented things from taking effect for a while."

Actually, that was Team Obama trying to get first decade spending down to under $1 trillion.

But your sad excuses hardly matter. You got your bill passed, and it has Obama and Pelosi's name on it.

The simplest way to reduce overall healthcare costs is simply not to provide healthcare to society's underclass.

Posted by: krazen1211 | November 3, 2010 7:12 PM | Report abuse

"That's because right wing opposition prevented things from taking effect for a while."

Actually, that was Team Obama trying to get first decade spending down to under $1 trillion.

But your sad excuses hardly matter. You got your bill passed, and it has Obama and Pelosi's name on it.

The simplest way to reduce overall healthcare costs is simply not to provide healthcare to society's underclass.

Posted by: krazen1211 | November 3, 2010 7:12 PM | Report abuse

In any case, the Democrats are about to take a bigger structural loss.

Republicans have a huge redistricting monopoly in a whole load of places, and in the one major place where Democrats have it (California), they can't use it.

That means guys like Joe Donnely who survived this year are likely to get booted in 2012.

Posted by: krazen1211 | November 3, 2010 7:20 PM | Report abuse

Just wondering if anybody at the Post realizes there are economists who are not university professors and not part of the administration. You know, people who work with real businesses, and who get fired it they make bad predictions.

I've never seen one interviewed or discussed in the pages of this column, but I hear they exist and I hope to live long enough to actually read about one!

Posted by: 54465446 | November 3, 2010 11:20 PM | Report abuse

My model didn't work. I wonder why the public didn't cooperate with my model?

Posted by: 54465446 | November 3, 2010 11:26 PM | Report abuse

Drum correctly used the model, using a guess for 3rd Quarter personal income growth. The BLS last week (well after Hibbs wrote the paper) published their 3Q estimate and, using it instead of the guess, the Hibbs model would have predicted 215 Dem seats (plus/minus 11). So it's worse than Drum thought.

Political scientists, along with all the political forecasters and pundits and statisticians, remain imperfect. That's okay. Drum and others should not, however, assume that any model or equation is correct as their starting point for interpreting Tuesday's results. Karl Smith touches upon this by suggesting the model should change in light of new data. More pointedly, the model may be quite wrong.

Oh, btw and to my (limited) knowledge unemployment has never been very useful for explaining or predicting election outcomes. The same cannot be said about other economic indicators that are more closely associated with voters' incomes. Perhaps that's worth keeping in mind this week.

Posted by: pjro | November 4, 2010 1:34 AM | Report abuse

Ummmm . . . that's not science?

You can't adjust a model after the fact of a failed prediction, rerun it, and then claim that the model predicts what happened in the past. (a) It isn't the same model. It is a different one. And (b) you can't predict the past.

What you've done is a bit of ad hoc numerology to make the model fit reality after it failed. That adjustment is not driven by an understanding of actual dynamics. It is simply doing whatever mathematical manipulations are necessary to force the model to fit. That's not science.

I'm not familiar with Hibbs' model but I am familiar with modeling data in physics. With enough free parameters, you can fit any model to any data. I frequently have this discussion with students when they claim that a tenth order polynomial is a better fit to linear data than a straight line because it goes through all the points.

When that happens, the telling issue is what happens when new data arrive. Does the model still fit pretty well, or do you have to readjust all your parameters to make it retroactively fit the new data point? If the latter, then your model has nothing to do with physical laws. It is just a random fit.

You also have to operationalize the meaning of the parameters you are using for the fit. In this case, presumably, the meaning has to do with the percentage weight each factor has in the decision process of a typical voter. Trouble is, the actual weights are probably variable -- some years the economy matters more than other years -- and in any case the process of deciding on a vote is probably nowhere near as orderly and well determined as the model implies.

Posted by: pj_camp | November 4, 2010 9:17 AM | Report abuse

pi camp wrote:

"Ummmm . . . that's not science?

You can't adjust a model after the fact of a failed prediction, rerun it, and then claim that the model predicts what happened in the past. (a) It isn't the same model. It is a different one. And (b) you can't predict the past."

Exactly!

However, what are the consequences of having a bad model when you're a university professor?

Posted by: 54465446 | November 4, 2010 10:36 AM | Report abuse

What a great idea. Attribute the worst electoral ass-kicking since 1932 to some esoteric mechanical model. How about this? The majority of voters are not persuaded that Obamacare is in their best interest, and the more Mr. Obama tries to sell it, the more they dislike and distrust it and him.

Posted by: DavidKeene1 | November 4, 2010 3:04 PM | Report abuse

I'm really just testing to see if Ezra's comment section is working, because it's not over at the Plum-line. Also testing my Troll Hunter script over here, to clear some stuff up. Sorry for any inconvenience.

Posted by: Kevin_Willis | November 5, 2010 4:46 PM | Report abuse

Post a Comment

We encourage users to analyze, comment on and even challenge washingtonpost.com's articles, blogs, reviews and multimedia features.

User reviews and comments that include profanity or personal attacks or other inappropriate comments or material will be removed from the site. Additionally, entries that are unsigned or contain "signatures" by someone other than the actual author will be removed. Finally, we will take steps to block users who violate any of our posting standards, terms of use or privacy policies or any other policies governing this site. Please review the full rules governing commentaries and discussions.




characters remaining

 
 
RSS Feed
Subscribe to The Post

© 2010 The Washington Post Company