Thursday, March 24, 2011

 

Is Your House (model) On Order?

Earlier this evening, Nate Silver posted a response to my House model. In a sentence, I agree with his basic finding (like many of his). In fact, I made it a week ago in my comment section. My model simply will not work for 1948. It has a difficult time for any year prior to 1948.

Issues such as those raised in Nate's piece about coding often plague political scientists. When does a new political era begin? What constitutes an aggressive war? I do not pretend to know the answers. I only hope to know them.

I have made my judgement about 1952 beginning a new political era that continues to this day. Nate clearly disagrees with that belief. My argument is pretty simple. World War II along with the Great Depression forever changed the political landscape of our country. Some might argue that then I should include 1948 in my model, as it is post World War II. My postulation is that World War II played a role in the 1948 election in ways we cannot model.

Truman won re-election despite having negative yearly growth in real disposable personal income per capita in the first three years of his term (before a recovery in the fourth year). While quarterly disposable personal income data was not published for the first half of Truman's term, my guess is that Douglas Hibbs' quarterly growth real disposable personal income per capita based Presidential model (which Nate and I agree is solid) would have difficulty handling the Truman election.

We know for instance that Ray Fair (who has his own economic Presidential model) had to make ad-hoc adjustments to the all the elections in which the President's term (including 1948) encompassed any part of America's involvement in World War I or World War II. My model is not immune to this World War problem, but considering we have not since had a World War and will probably not have one for a long while, I am not particularly worried for future forecasting.

Another issue that Nate finds fault with is my use of the war variable, which my dataset codes as true for 1976 and 2008. As I noted in my piece, this variable is merely a dummification (not a word, I know) of Douglas Hibbs' Fatality variable. I apply the dummy only to years in which the Majority party in the House differs from the party that controls the Presidency. To answer Nate's question, Libya will count as a war in my book if Douglas Hibbs counts it in his.

A fair question to ask is why does this variable not apply to years in which the party in the White House is the same as the Majority party in the House. Note that including it does not affect my findings. Why? The answer is there is another variable that encompasses the war variable in years in which the President's party is the same of the Majority party.

As Hibbs' model illustrates, Presidential vote (which is my model's main variable for predicting House seats in years in which the Majority party is the same as the President's) is mostly a function of the quarterly real disposable personal income per capita growth and military fatalities. Thus, it would make sense (to me anyway) that the model warps out any effect my war variable would have when the Presidential vote variable is in effect. The question that should be asked is why does economic growth not matter in House elections in which the Majority party differs from that of the President.

Of course, coding of wars is always interesting to deliberate. Hibbs does not count any of the fatalities under Nixon's first term (and Obama's first term for Afghanistan and Iraq) against him because he "inherited" the war from Johnson. I happen to agree with Hibbs, but I know some who do not. The point is that a "right" answer in coding is not always apparent.

I also want to address a few other points.

First, Nate and I have "known" each other (at least in the blogosphere) for about a year now. During that time, I have felt (and continue to feel) that we have had a respectful and amicable relationship. I have "critiqued" some of his work (see here), and he has returned the favor. This exchange is healthy for me (as I hope it is for Nate), and more than that, it is important for the reader. Without this back and forth, anyone could post a poor model at a whim, and we would be none-the-wiser.

Second, part of the reason that it was so easy for Nate (or anyone) to check my work is because I post my datasets for everything I do online. If I do not, I tell readers (and I really do mean it) that they can email me. I wish more online political writers would post their datasets. It is the right thing to do.

Third, margins of error can be tricky for any out-of-data forecast. I do not pretend to know that my +/- 10 will definitely turn out to be right (e.g. it might actually be closer to +/-12), but it is based on something concrete (the existing dataset). I would rather make my margin of error that way then create some large arbitrary error to cover my behind. Some can point to this simple Gallup house model as an example where the margin of error (+/- 11 seats) of the in-dataset did not quite cover the spread in 2010 because it ended up being off by 13 seats. Unlike that dataset however, my model is not reliant on some measure (a Gallup poll) that is subject to additional error.

Finally, a wise man once said, "to be clear, accuracy ought to be the paradigm here. We're not trying to prove or disprove anything to an academic certain degree of certitude; we're trying to make a forecast... [My] model may be wrong, but I'd rather fail by being too ambitious than too stubborn."

I agree with him.

EDIT: One more point on degrees of freedom and number of cases. If you take the full model, you have 6 variables on 15 cases. If you take the not full model (i.e. only having the interaction variables without the original), you have 4 variables on 15 cases. I'm in nowhere comparing model to Douglas Hibbs' midterm House model, but there you had 3 variables on 15 cases. The Gallup model discussed above also had 3 on 15.

If you were to breakdown my model to the years in which the Majority party was different the President's party (which is basically my argument... that we have two types of House election is Presidential years since 1952), you would get the same result with 10 observations but only 2 variables (war and previous seat count). I have seen some suggestions that a solid model should have at least 10 to 20 times as many observations as variables. Most academic Congressional models and mine fail this test.

The real question, however, is whether statistically the model is too "tight" a fit. I have made the point that 1952 began a new political era. This belief has worked through 2008. In fact, as you can see in the comment section of my original post, the model did just fine in every year since 1952. If 2012 ends up being different from the '52-'08 dataset, the model will meet Reaper, Grim.

However, I really do not believe 12 will mean the model's death. Good aggregate based modeling (such as Hibbs and even the Gallup model that Nate does not like) did just fine in 2010. Most of them were inside their margin of error, and even the Gallup model missed the margin of error by just a few seats (or less than three standard errors). I will be making a post that the generic ballot (at least at this point) also supports my model.

The true test will be in a year and half. If it is clear my model is going down the drain before then, I will jump off the bus faster than fans of the Yankees when they are heading for last. And as those who know me can attest, I will come back here and admit my fault.

Comments:
I think Nate's main point was you have 6 degrees of freedom on a fit for 15 data points. The brittleness with respect to 1948 is evidence for this criticism, but it's not the core point.
 
What about Nates critique about "overfitting"? i.e., the ratio of variables to cases.
 
What kind of "roll" did it play? Not a Kaiser Roll, he was out of power long before.
 
Corrected to role. Thanks :).
 
Couple of things:
1)You have a lot of 'beliefs'- that 1948 is a justified dividing line, that the Fatalities variable is properly applied (and that eg the invasion of Panama, the invasion of Grenada, Beirut, don't count for this).
And I can't help wondering if you believe these things because they produce a good fit for the model. eg you pretty much explicitly say this about 1948- you believe that 1952 is a new paradigm because 1948 doesn't fit the model.
2)Any model that has special variables that apply only to one data point (and I think that the use of the fatalities rule, with it's "or" divider, is actually just two such variables, one for 2008 and one for 1976) is suspect. It allows an ad hoc adjustment to the model for specific cases. That is not really modeling anymore, it's hand-picking values for specific data points.
3)Are you willing to take a bet based on your model? If Nate's math is correct in his column, you've given 100,000 to 1 odds against a Dem takeover. Would you accept a cash bet at 10,000 to 1? 1,000 to 1?
 
sry, one other point:
You said that you modeled the 2008 election using this method based on the data up to 2004. But it is only after the 2004 election that the model would've been modified, changing the fatality variable (adding language to include 2008's scenario, where it would've only had the 1976 language previously, unless the writer was prescient). That is, to my reading you've basically added a special variable (via "or") to account for 2008 post-facto and then claimed that the model does a good job of predicting 2008 based on the data through 2004 *including* a variable that could only have been extrapolated after the election.

So, my expectation for your model is this- it could be right, that result is certainly a possibility. But if it's wrong, you'll add another variable (perhaps via another "or" clause) to account for the discrepancy. Perhaps Libya will become a war. Or perhaps the variable can be tweaked so that Nixon's inheriting of Vietnam doesn't count but Obama's inheriting of Iraq & Afghanistan does. Perhaps the model gets an entirely new variable for economic crises.
 
Karmon,
Thanks for the comment. Let me hit on your points.

1. I'm not making the war variable. I really am not. As I said, I took Douglas Hibbs' fatality variable and merely made it into a dummy. Wherever his variable for fatalities is greater than 0, I put in a 1 for years in which the President's party is different than the Majority party in the House. I did not make it. I would not do that. That would be far too arbitrary on my part.
2. I guess that is one way to look at '48. If it were just my model, then I would agree entirely with your point. The fact is that it is not just my model. 48 is a weird election for a number of reasons. As I mentioned the economic models also have great difficulty in those years. The national vote to seat curve is another thing that is way off. World War II was still having an impact on election in ways that I do not believe anyone can model. The '46 election suffered from many of the same problems (see Bafumi et. al's generic ballot 300 days out model).
3. If you want to put in an extra buffer zone of error, that is your prerogative. Nate certainly does, though Nate has made it a habit to do so (see his +/-40 seat cone of error on the eve of the '10 House election).
4. I do not what I would have done. I can tell you that I was working with Presidential modeling using fundamental factors when I plugged House model into the computer. Thus, I was also working with a war variable. The model would have done just fine (as the piece notes) with 2008 out of dataset forecast.
5 (which may be out of order with 4). We'll have to disagree on cherry-picking. Again, I'm not making the variable. I certainly did not come to it that way. It wasn't like "oh hey this didn't work, let's try this". Unfortunately, we only have so many elections to use it. I can only say that it worked in 08, despite only having one case before. Interestingly, Nate has done ad-hoc adjustments such as for the Maine gay marriage vote when there was only a few cases of off-year elections.
6. The bottom line for me is I see a clear line dividing Prez years with the same party controlling the House and Presidency and those in which they don't.
7. I don't gamble, not that amount. Nor do I even have that type of money (I'm a student). I'd go 10 dollars to a cent or 10 dollars to the 10th of a cent though.
8. I do not believe your reading of the model is right re: 2004 modification. For one thing '04, does not use the war variable because it is reliant on the Presidential vote because the majority party in the House is the same as the President's party. Clarification on your point would be helpful.
9. Your expectation is wrong. I'm not making the war variable. If the model is right, I throw in the burner and say Nate was right. I said that up above and stand by it.

P.S. If you email me, I usually get to the comments quicker. Feel free to post them in the comment section, if you want it to remain on the public record.
 
Thx for the reply- forum is fine, Im not in a particular hurry & maybe others will have something interesting to say.

1)afaict Hibbs is using the KIAs quantatively, so eg Panama wouldn't have much of an impact, it'd be lost in the error. Since you've dummied this into a boolean, you're put to make the decision: does this count as a real war?
And it's certainly not "greater than zero [casualties]"- there were casualties in eg the invasion of Grenada, and the Dems controlled the House in 1984. There were casualties during the invasion of Panama.

3)"If you want to put in an extra buffer zone of error, that is your prerogative. Nate certainly does..." Well, Id hope that the buffer zone represents (or tries to represent) the actual expected deviation. Yours represents the model output, and that's fine- but we have to put the caveat on this that the model is (IMO, anyway) far from perfect, with only 15 data points and a large number of variables to explain them.
And I don't think you're giving those sorts of caveats. If the economy craters, there's nothing in model for that. If a smuggled NKorean nuke takes out LA in Oct '12, there's nothing in the model for that. Because the model has been trained on a very small dataset and the universe of possible election-impacting events is huge.
So I don't think the confidence in "it looks like it will still be Speaker Boehner come January 2013" is justified.
That's also the point of the bet. I dont know if you're a basketball fan- at the start of the season the Atlanta Hawks were about 3000:1 against to win the NBA championship. You seem to be arguing that we should see that about 30x as often as a Dem majority in 2012. That seems pretty far off, probability-wise.
If Nate had error bars of +/-40, maybe he's cautious, or maybe that's all he could get out of the data- there's as much fault in being overconfident as in being overcautious.
(To put it a third way on the bet- if you really think this, you should feel pretty confident borrowing every penny you can and betting it on the 2012 House election; you'd be basically getting free money from folks offering 2:1 odds of a GOP hold.
If you argue that the 100,000:1 odds could actually happen & you'd lose your tuition etc, that's true- but you're much more likely to get killed in a dozen different ways this year, and you probably don't worry much about any of them.
If someone gave me 1:2 odds on the Hawks winning the NBA championship I would bet everything that I could get my hands on against it.)
 
I suppose Im marveling at this contrary-to-common-sense claim made by the model: the electorate favored the Dems in 2008 because they won the House in 2006. Not because the Iraq War had dragged on long enough to be viewed as a bad decision and a quagmire. In fact, the model predicts that if the GOP had held the House in 2006, they wouldn't have suffered in 2008 either. Or, contrawise, if the war had been started in 2005 rather than 2003, no effect.
That seems to me to be a limitation of working with such a small data set. Again, the universe of possible events is too large, and the danger of a correlation between 'opposite-party control' and 'punishing the WH for a war' is too high when there are only a few data points to work with.
The point about 2004 is this: if you had created this model in 2004, prior to adding in the war-but-only-if-the-WH-and-House-are-split variable, you might have accounted for the 1976 case with a resignation variable rather than a war variable- a pretty significant event, after all! After 2008, you can see an opportunity to account for both cases with one (oddly-worded) variable. But the former, more plausible variable wouldn't have predicted the 2008 House election.

I do find it interesting to see results such as presidential share not impacting change in Congress. It's an interesting model. But again, I wouldn't put my money on a 100,000:1 odds against a Dem takeover, and I suspect you won't either.
 
On the prelude to the first point, I hope the one thing that can be taken from this whole thing is I (anyway) try not to take any of this stuff too seriously. More than that, I try to explain my reasoning in a clear to understand manner, and I'm always open to suggestion.

On point 1, yes I know he is using a continuous variable there. When you use a continuous variable in my model, the result is rather interesting. It does work to explain past results well, but it would have really screwed up in 2008 (i.e. said the Democrats would control 81% of the House). The re-fitted model with fatalities works well however as it corrects for the new data. Perhaps, that is a point to your argument about out of dataset surprises? IDK.

You'd have to ask Douglas Hibbs about Grenada. I believe he really is counting all the wars with fatalities. There were few fatalities heading into the 64 election where he has them listed and few in the second Nixon term, and he has them listed. He doesn't count the first Gulf War either, and there were enough fatalities there to at least make a decimal listing. So I do not believe it is a matter of a continuous variable. That said, I perhaps should re-name my variable "anywhere where Hibbs has a fatality number". Not sure that is any better.

i'll admit I could have been clearer, and I think I was in the comment section of the first post. Specifically, about saying how it does not work to explain anything before 52 AND re-running all data out of forecast. I would say that is the frustrating thing here. I think many people (including Nate in some situations) do not include any sort of hardcore number for error. When he does, he has a history of also getting "had" by trying to apply MoE's from past regression to future results (http://www.pollster.com/blogs/regression_analysis_of_the_dem.php?nr=1). So, it is a little weird to get this lesson from him. I show all the work, so at least people can agree or disagree. I will say the model isn't far from perfect in explaining anything post-52... It's done a good job of that.
 
On the economy cratering, I think that is where we differ on that one. My argument would be that wouldn't matter. Indeed, track the generic ballot in 08. The thing really didn't move the entire year. I don't believe the economy had that big of an impact on voting patterns. I know Bob Erikson would certainly argue that. As for NKorea, I got no clue. Perhaps, I could have said "if everything goes on without anything crazy like North Korea". I almost wish it did happen though, so we could test it. Certainly, there have been strange events in the last 60 years, which haven't done anything to the model.

On the Speaker Boehner (gosh I can never spell that right), I think I have more confidence than you do. I think it will be him. I am that confident of the result and that is what this is all about (to me anyway). Baseball or football would have been better. The last memory of the Hawks were them losing in 4 straight (I think or maybe it was 4-1) in 1999 to the Knicks in the second round. But to your point, probability wise, I think you're probably right. I think it is 95 times out of 100 that the model gets it right, but those 5 times it could be way off. I obviously know that is not the way probability works. That said, I feel far more confident than 7:3.

I cannot speak to Nate's confidence, you'd have to ask him. Here is what I will say though about my mindset in all of this. I know that elections tend to be stable. Especially when looking at a large aggregate body like the House. The generic ballot stayed pretty darn consistent in 08 and 10 when keeping the population (RV and LV consistent). Right now, we're probably looking at a Republican edge of a few points (somewhere between PPP and Rasmussen and right near Dem Corp) on the generic. While vote seat curves can be off that few points would put the final result well within my MoE. It'll be interesting to watch. Overall, I do look at these other factors before going blind into the wind. Over anything is bad (I gotta lay off the Lucky Charms, for instance).
 
I will say re: money I'd of made a ton of change in the last election and 08. Maybe I should be betting. As for dying, statistics are good on a computer. The mind, unfortunately, is not always as reason based.


I don't believe the model is saying that about the dems winning in 06 and thus 08. It's saying that they were likely to stay in power because they already were in power. The model has nothing in it about prior seat gain, only prior amount of seats held. The model says they were to gain seats because Bush had started a war with fatalities. Neither of these claims seem too far-fetched to me. As for the Republicans holding the house, again that is a misread of the model. In that case, the model is saying that the Republican seat count would have been predicted by the presidential vote share of McCain (as the Republican share in 04 was dictated by Bush's share or Democrats' share was dictated by Johnson's in 68). That I think is what is most interesting about the model. It is saying prez coattails don't exist in years where the Majority party is different than the party in the White House. That actually makes a lot of sense to me based on results over the last 60 years. Per war starting in 05 vs. 03, that's a misread. Bush started the war so it's his party war regardless of 03 or 05. It goes down as a true for the 08 election.

As for your argument about the "universe is so large" and the model cannot control for it, more than fair, but we're going to have to see. I'm going to ride this sucker out until the tires give. I see no reason they will.

As for the resignation variable, I again cannot answer a hypothetical. I was working with Hibbs' model, and I stumbled upon this finding. I actually worked backward. First plugging in a war variable for all the data and then seeing it only worked for years in which the parties differed. I am not Ray Fair and won't ad-hoc it. I certainly did not here (at least in the steps I took). And yes, the variable is probably worded poorly. That's my fault, not the model's.

In closing, the thing that is most frustrating about this whole thing. If we take out my quip about MoE (or better yet, clarify it), then we can concentrate on the result. The result is what is important (to me anyway).
 
1948 wasn't a WWII issue. The parties were perceived differently in terms of the economy. Gallup polling showed a 20-point advantage for Democrats "if hard times returned." Hard times did return, and they swung toward the Democrats. Unsurprisingly, they did not want the party of Hoover governing if the Great Depression was reemerging. Several contemporaries noted this as the reason that the vote swung, e.g., Samuel Lubell.
 
As a statistician who teaches undergrads, I have to say, without any malice, that your confidence in the predictive accuracy of your regression strikes me as a classic move by an inexperienced practitioner without deep theoretical training.

Basically any regression on social data with fifteen observations and six predictors is very likely over fit, and presumably you considered some fairly large number of other variables and interactions? That diminishes the confidence you should have in your predictive accuracy even more.
 
Per war starting in 05 vs. 03, that's a misread. Bush started the war so it's his party war regardless of 03 or 05. It goes down as a true for the 08 election.

Ok, but that isn't justifiable as an inclusion in the model. That is, you can make that your variable, but there's no justification for it compared to the one I misinterpreted. There is exactly one case here, and generalizing from one case to make a model is guessing, not modeling.

Perhaps, I could have said "if everything goes on without anything crazy like North Korea".

That's the point about all of the possible events that haven't happened. Or, I suspect, happened and had their effects mistakenly assigned to another cause (ie Nixon's resignation). There's a huge universe of possible events, and it's not at all unlikely that one of them would happen. Again, if Nate's extrapolation was correct and your model is predicting 100,000:1 odds of a Dem takeover- then you've got the distribution wrong, bc I suspect that many of those unlikely events are more likely than this. That's get-struck-by-lightning territory, oddswise.

The model says they were to gain seats because Bush had started a war with fatalities. Neither of these claims seem too far-fetched to me.

Well, the model says that they were to gain seats because Bush started a war with fatalities *and* the Dems controlled the House. Let's posit that, if a war becomes unpopular, then it gradually does so- or, at least, that it does not do so in response to election cycles. Given the tiny data set, it's not unlikely that the response is first seen in the mid-terms in the *one* case where the House was switching hands before an election where this variable was true. That is, the war causes both the House to flip and the continued degredation of the GOP position in '08. Not that the House flip causes the GOP degredation in '08.
You might say that causality here isn't important, but it is insofar as you've got this limited data set. If Bush starts the war in '05, it seems entirely possible to me that the House doesn't flip in '06. Then, your variable is off for the '08 election, because you've made a result (the House flip in '06) a cause of the '08 election, rather than just another result of the war dissatisfaction.
In that alternate reality, I suspect that you would slightly alter the variable so that, if a President *started* a war, it would count regardless of who controlled the House.

I actually worked backward. First plugging in a war variable for all the data and then seeing it only worked for years in which the parties differed.

You've generalized this model from *2* cases where the variable applies. And in at least one of those cases, I suspect that the variable isn't actually the cause- and you've had to ignore one of the largest political events in the post-war era. And you've had to pick and choose which wars count and which ones don't afaict.
Generalizing from, say, a thousand cases is one thing- the danger that there are 10 or 100 or 1000 variables that you've missed that expain the effect you're seeing is smaller. Here, what are the odds that you've missed *2* things, and mistakenly thought that they were one thing, based on a (again IMO) dubious-sounding variable definition? Especially, again, considering that we've actually got a good candidate for one of the things (ie Nixon's resignation in disgrace).
(Ive heard that exit polls in '76 showed that Nixon's actions and Ford's pardon weighted heavily on voters, and Ive seen charts of voter identification for the GOP dropping precipitiously after Watergate- as opposed to gradually degrading in response to the Vietnam War).
 
And yes, the variable is probably worded poorly. That's my fault, not the model's.

I think it's the model's- it requires the variable to catch specific cases and exclude other specific cases. Usually (not always) a good variable should be easy to label, like "delta in unemployment" or "inflation rate". This one doesn't submit to that sort of treatment. If it were big and complex, and explained a lot of cases, that would be one thing. That it's big and complicated and explains 2 things, I think that's a bad sign.

On point 1, yes I know he is using a continuous variable there. When you use a continuous variable in my model, the result is rather interesting. It does work to explain past results well, but it would have really screwed up in 2008 (i.e. said the Democrats would control 81% of the House). The re-fitted model with fatalities works well however as it corrects for the new data. Perhaps, that is a point to your argument about out of dataset surprises? IDK.

I guess yeah- if you made this model up in '05, then '08 would've taken you by surprise, as you say. Now, having accounted for '08 with some changes to that model, you may (or may not) be surprised by '12.
And it still isn't clear to me how you're excluding Panama, Grenada, etc. The only way I see to do that is to put in an arbitrary threshold of KIAs or length of time the US is engaged etc, and that would once again be tailoring the variable to fit the past results. Which would continue to erode my confidence that it's really predictive.

That said, I perhaps should re-name my variable "anywhere where Hibbs has a fatality number". Not sure that is any better.

That's really no good- I think you need to own the input here. It's not like the KIAs for various military adventures aren't available. Treating this as a black box because it happens to work as a feed for a variable that explains two cases is convenient, but unjustifiable in modeling terms. And you should certainly change the claim of a zero-KIA threshold, since that's not the case.
 
Sean,

See Brendan Nyhan. He would disagree on your view. In fact, he believes the Dems won because of a good economy due to late growth. I'd say the issue is far from settled on that score.
 
GoodEpic,

My profs would disagree as would most who know me. I probably mis-worded MoE in this situation, but your statement is far too broad.

Oh and btw, Nate was doing the exact same thing in 2008 with errors falling all around... But nobody seemed to then.

Is he not classically trained enough?
 
Oh and GoodEpic, I believe I addressed your concerns in my edit section of the last post.
 
Kamron,
I'm glad we can continue this discussion in a respectful way.

As per your first point, I'm not really sure modeling vs. guessing is a great distinction. Obviously the first time you have any case before a second case, you are going to have to make a decision on what should and should be put into the model. For instance, Douglas Hibbs may be using a continuous variable, but he noted in his earlier pieces (I believe 2000 specifically) that those continuous variables could be replaced with two 1's in 1952 and 1968, and the model would work just as well. He made the "call". I, of course, did not have to make the "call" for 2008 before the election occurred. What would I have done? I have zero clue.

As lightning strikes, I think I could have improved the wording of my MoE to make it clear that I was not saying a be all end all. That if no earthquakes occur etc. If that is really what this whole thing is all about, then I gladly concede the point. My wording was poor.

As for the House not flipping in 06 then 08, I am not sure why this is a problem. The results then in 08 would have been associated with the president's vote. That vote as I have been saying is mostly explained by the economy and war fatalities. As for the rest, I cannot answer hypothetical questions. I can say, however, that I did not go about looking to alter variables beyond recognition. I can also say that the '76 case would not be true because Nixon nor Ford started the war. If Hibbs has anything but a zero, I put down a 1.

As for the generalizing, I think many models have to do that. I wish we had a prez election every year where we could test the model. I do agree that I could be "mistakenly" assigning a variable. I do not know, and I cannot imagine that anyone, unless they have a crystal ball, knows either. As per 1976 and Nixon, it is an interesting question. If Watergate was playing such a big role in the presidential contest or congressional contest, then why does Hibbs' midterm Congressional model for '74 and Presidential model for '76 do such a great job of predicting the vote? Is that spurious correlation? Again we only have 2 variables on 15 observations (or 2 variables on 12 when his model was first published in 2000) for the prez model, and we only have 3 variables on 15 observations for his Congressional model.

I think we'll have to disagree on the war wording. I, unfortunately, am not a great communicator in the way I wrote it. It's a variable that captures war fatalities in years in which the prez party differs from the majority party. Is that better? The rest of the clause is something that not even Hibbs could fit into a nice sentence.

As for Grenada, I am not sure I can be any clearer. I am just going going by the Hibbs' dataset. Now maybe it is on me to then go in manually and re-adjust for Grenada and then for the Persian Gulf War (which Hibbs also does not have). I found in my initial looking at the Hibbs dataset that he does not include fatalities for a number of smaller conflicts (even those where the number of fatalities matches those for say '64 Vietnam). I do not think I have to re-define because I already am saying this in my initial post " A dummy (1 or 0) that is 1 when the party in control of the White House is different from the majority party in the House, and the President has started or maintained for more than one term an "unprovoked, hostile deployment of American armed forces in foreign conflict" (as defined by Douglas Hibbs) which has resulted in at least 1 fatality during the past term." Here I clearly say where I am doing.

Oh and I meant to say that you can feel free to respond to in public, but also send me an email, if you would not mind. There are a few personal questions I'd like to ask.
 
This was a very good respomse. I really thought Nates critique was damaging but I think you've explained yourself well. I do think your margins of error are too tight but as you say, that's the way the math turned out. James honaker is working on a paper that takes the uncertainty from things like polls and carries it into the std error of coefficients but I guess you didn't use polls (I didn't read your original article) so I guess that wouldn't help
 
Post a Comment

Subscribe to Post Comments [Atom]





<< Home

This page is powered by Blogger. Isn't yours?

Subscribe to Posts [Atom]