Comments on FiveThirtyEight: Politics Done Right: On Moon Landings, Michelle Malkin, P-Values, the C...

酒店經紀人, 菲梵酒店經紀, 酒店經紀, 禮服酒店上班, 酒店小姐兼職, 便服酒店經紀, ...

2009-10-10T08:10:33.803-04:00

酒店經紀人,
菲梵酒店經紀,
酒店經紀,
禮服酒店上班,
酒店小姐兼職,
便服酒店經紀,
酒店打工經紀,
制服酒店工作,
專業酒店經紀,
合法酒店經紀,
酒店暑假打工,
酒店寒假打工,
酒店經紀人,
菲梵酒店經紀,
酒店經紀,
禮服酒店上班,
酒店經紀人,
菲梵酒店經紀,
酒店經紀,
禮服酒店上班,
酒店小姐兼職,
便服酒店工作,
酒店打工經紀,
制服酒店經紀,
專業酒店經紀,
合法酒店經紀,
酒店暑假打工,
酒店寒假打工,
酒店經紀人,
菲梵酒店經紀,
酒店經紀,
禮服酒店上班,
酒店小姐兼職,
便服酒店工作,
酒店打工經紀,
制服酒店經紀,
酒店經紀,
菲
梵,

I'm sure no one's reading this comment thread anym...

2009-06-11T11:24:40.297-04:00

I'm sure no one's reading this comment thread anymore, but what the hell.

The company I work for does a fair amount of business with car dealers. When the Chrysler/Dodge dealership closures happened, we were a bit puzzled by how the decision played out in our local area. Apparently, the company has some sort of rating system for dealerships, and in our neighborhood there were some higher-rated dealerships closed while lower-rated ones stayed open. Why?

Well, my boss reports that he was at some meeting talking about this with a local car dealer, and the dealer opined that it was pretty clear to him how the decision had been made: all the guys who lost their dealerships were jerks and no one liked them!

There's a serious consideration hidden inside this anecdote: it's possible that the list of losers wasn't even compiled at the high levels of the company, but may have been farmed out to regional/local offices to decide according to their own preference.

All in all, this does not preclude the possibility...

2009-06-09T22:44:52.691-04:00

All in all, this does not preclude the possibility of their having been some favorability for large Democratic donors. I don't believe this to be the case. But without knowing the criteria we're just left with our biases and guesses. That's what a lack of transparency gets you.

Also, this shows how much of a mess it is to have politicians making decisions normally left to the free market. Feelings will be bruised and people will be angered. If there's a lesson for you democrats to learn, that'd be it.

Many of the same people here were less than a year...

2009-06-09T22:24:24.867-04:00

Many of the same people here were less than a year ago grasping for straws and cared nothing for significant P values. Voter dis-enfranchisement and republican conspiracies were a favorite topic of discussion. Who needs statistical proof for that, we all know it's true! The 2000 election was stolen and exhibited systemic efforts to stop minority voters from voting, based on anecdotal evidence with any kind of statistical analysis a distant concern if performed at all.

What a joke. When it comes to discrediting Republicans, most of the people here are very liable to believe it and don't need statistical proof. But to throw your own guy under the bus you demand more than anecdotal evidence since it runs counter to your view of the world. Everyone has a high standard when it comes to overturning a worldview they believe to be based on a lot of evidence, and a low burden of proof for things which confirm already known "facts."

So sorry, righteous indignation from people on the left bemoaning conservatives low standards of proof rings very hallow following the last eight years.

The coin-flipping example is a good one, but there...

2009-06-03T15:29:24.732-04:00

The coin-flipping example is a good one, but there's more to it than Boonton's example suggests. Suppose you take a quarter that you have just received in change, flip it 7 times, and get 7 heads. This is a very unlikely occurrence (less than one chance in a hundred, or p <.01) IF it's a fair coin (one that has an equal likelihood of coming up heads or tails--the null hypothesis). However, it's still more likely that the coin is fair and that this particular result is due to chance than it is that some other explanation is involved (for example, that the coin is biased, or that someone is psychically manipulating the coin). In this case, the “chance variation” explanation is unlikely, but all other explanations are even more unlikely.

This illustrates an essential point of scientific method--that you want to test the plausibility of your proposed explanation relative to ALL other plausible explanations. From this point of view, you have to consider the likelihood of possible explanations for the result BESIDES chance variation in the sample. Lots of people have posted plausible explanations for the apparent association between dealers' campaign donations and dealership closings, in addition to a deliberate closing of Republican dealerships. In addition, Anandakos made the extremely important observation that the likelihood that anyone could slip through a deliberate closing of Republican dealers without someone blowing the whistle is extremely low.

However, statistical significance tests (p values) only assess ONE possible alternative explanation--chance variability in the sample. They say NOTHING about all the other possible explanations for the result, or how likely these are relative to the "conspiracy" explanation. Such tests are useful only if chance variation is the major alternative to your proposed hypothesis.

I am not schooled enough in statistics to have a v...

2009-06-03T14:07:01.827-04:00

I am not schooled enough in statistics to have a valid opinion about the analyses presented here.

However, I'd like to point out that DirectorBlue and other conspiracy supporters have one enormous hill to climb: that is that SOMEONE in Chrysler's upper management just about has to be a rabid Republican. There is simple no way that such a conspiracy could be kept quiet.

Since this has been boiling in the blogosphere and even in the larger media for at least 48 hours now, the likelihood that the theory is true is shrinking rapidly.

Can anyone here entertain for one microsecond the idea that a partisan Republican having knowledge of such an order from the "car czar" would not have blast-faxed it to every media outlet this side of Saturn's moon Enceladus by this time?

I think this might be a simplier way to explain He...

2009-06-03T13:28:47.035-04:00

I think this might be a simplier way to explain Hedge's p-score problem.

Say I think a coin is biased. I flip it twice and it comes up with heads both times. Biased?

Well clearly if the coin was biased towards heads, getting heads both times is consistent with that. But there's a 25% chance that an unbiased coin will yield head-head when flipped twice. The 'p-score' in that case is 25% or we could say our confidence level is only 75%.

Well what about the 12.5% for the relationship between Clinton donors and not being closed? Well if I said a coin was biased based on only two flips, you'd laught at me. Well suppose I only flipped it 3 times and got three heads? Guess what, the odds of that are 12.5%!

When Keith had a report on this poppycock, right-w...

2009-06-03T01:37:41.538-04:00

When Keith had a report on this poppycock, right-wing conspiracy, he mentioned that you found 88% of auto dealers gave to the GOOPers, but also stated (paraphrasing) 'Another [unnamed] blogger found a 92% correlation between being an auto dealer and contributing to the GOOPers.'

The best part was the republican congressman who owns a dealership slated for closing complaining, and he voted against the auto bailout.

What did the moron think would happen to his business if the auto makers collapsed.?

I am a "geostatistician" and will archive and incl...

2009-06-02T21:42:11.503-04:00

I am a "geostatistician" and will archive and include this thread for my class -- it ain't geostat, but it sure is hypothesis testing. And it also confirms me in my Bayesian leanings....

An excellent analysis that closes the door on the ...

2009-06-02T18:22:22.390-04:00

An excellent analysis that closes the door on the subject (and I say that as someone who was initially concerned about the reports). Good work.

Very well analyzed and explained even better. Than...

2009-06-02T16:37:57.728-04:00

Very well analyzed and explained even better. Thank you Nate.

And even if the findings were statistically signif...

2009-06-02T08:31:08.603-04:00

And even if the findings were statistically significant, that only shows correlation, not causation. OTOH, in politics correlation is sufficient to run campaign ads and smear campaigns.

Does no one remember the 2008 primaries, wasn’t Ru...

2009-06-02T01:45:28.490-04:00

Does no one remember the 2008 primaries, wasn’t Rush exorting people to cross party lines and vote for Hillary…could it be the ever so slight, not very significant Hillary leaning of the dealers was do to some Repub dealers contributing to Hillary of the long grueling Dem primary in the hopes of dividing Dems…so while Dem dealers were likely split between Hill and Obama during 2008, Repubs would contribute to Hill…do not know if dealers could only be labeled as being supporter of one candidate or of any and all they contributed were counter in this…if any and all contributions were counted…Repub contributions to Hill seems simplest explanation…please correct me if this is an incorrect possiblity

I popped in to complain loudly about the "95 perce...

2009-06-01T23:31:48.652-04:00

I popped in to complain loudly about the "95 percent chance of the hypothesis being true" line (which is absolutely NOT the case)... but I see a few people have already picked up on it, so I won't labor the point, other than to say it's a major bugbear with students and professionals I encounter; I find myself having to correct this particular misunderstanding quite frequently.

So they run 6 hypo's, and find the most unlikely s...

2009-06-01T15:18:48.439-04:00

So they run 6 hypo's, and find the most unlikely set would occur, by random chance, 12.5% of the time.

So lets say I had a sack of marbles, made some blue, red, and non-colored. Then I labeled some of the blue "O" and "C", and some of the red "Mc".

Then I pull out 25% truly and completely randomly (I'm blindfolded, or don't look into the sack, etc).

Then I compare the blue inside the bag to out, along with the red, then the "O", then the "C", then the "Mc", and even the noncolored against the ones I labeled...

In other words, as Nate said, 6 hypos at once.

Wouldn't I on average get a 89% correlation or better, as a high water mark, using pure randomness?

(100 - 100x.89^6) = 50.3

(Obviously this ignores that things can be 2 things at once, which would allow for extra opportunities for finding 2 sets of "proof" at once, i.e. both blue and "O" marbles have been "saved", or both red and "Mc" marbles were "targeted").

So not only does this show nothing, it shows the data is actually WORSE THAN YOU WOULD GET ON AVERAGE if you tried to run 6 sets of data on pure randomness, and labeled the least likely result "proof"...

So if this isn't enough to get them to quit on this one and wait for the next opportunity to shout conspiracy, apparently nothing will.

-----

As an aside:

Using 6 unrelated data sets, you actually have over a 1/4 (26.5%)shot at reaching the magical 95% correlation or better with 1 of the 6:

(100 - (100x.95^6)= 26.5

You need 14 hypos at once to increase the odds of crossing .05 being more likely than not:

(100 - 100x.95^14) = 51.23.

Looks like they got a little unlucky, really.

Especially considering all the reasons they might have had better than just random odds, i.e if more rural dealerships closed than urban, or more gluts of domestic car dealers occuring in "red" towns in the US vs "blue" towns.

What would be more interesting would be to look at...

2009-06-01T14:24:51.841-04:00

What would be more interesting would be to look at geographic distribution of dealerships being closed versus those being left open. Dealerships are local markets, and there are pretty big efficiencies of scale/scope involved. Looking at the data on a national level might obscure leaving dealerships in market positions that would be favorable to one group but disfavorable to another group. IF there were a dealergate, I'd expect it to play out on a local level rather than a nationwide one.

I am a faithful reader of zero hedge. Looking at the chart, I simply couldn't explain how Ms. Singer came to the conclusions that she did. What I thought her data showed more was the human likelihood to not ax people in the back whom one knows personally. If the car czar is married to the head of Clinton's campaign finance committee, chances are that he met at least a few dealers who were big donors to the Clinton campaign. If there were 53 dealers who donated to her campaign total, that's a small universe - just a couple such instances would make a difference in P. I'm more likely to favor/trust people that I know vs. faceless car salesmen.

With that said, anyone who patronizes one word of Michelle Malkin's is a rube. She's kind of highly negatively correlated with rational statements and truth.

@directorblue: Why was the DealMakers auto group ...

2009-06-01T14:11:38.573-04:00

@directorblue:

Why was the DealMakers auto group protected from closing despite its teetering on the precipice of financial and legal collapse?DB, anybody with experience with this kind of decision-making will tell you that whenever you have a high ratio of population to variables (in this case, a population of 3500 dealers to...maybe 10, 15 variables? That's just a guess.), a few will inevitably slip through the cracks. I'll bet a leftwing nutjob could find Republican donors that fit a similar description as you gave above.

[I]n some cases, those parties are billionaires who do not want their names listed as primary owners in these types of documents.And you found one example. There are probably Republican donors for which this is true as well. It's YOUR responsibility to prove a hypothesis that you've proffered, and show that if we update the dealers to include ALL owners (don't forget to update Republican owners also!), then the relationships become significant.

You're demanding that the Administration (or someone acting in proxy, such as Nate) prove a negative (that the closures were not biased), which is a logical fallacy known as an Argument From Ignorance:

http://en.wikipedia.org/wiki/Argument_from_ignorance

@ronebofh:

NICE Hot Shots reference. I approve.

@Buckeye_Tom:

Your hypothesis cannot be proven or disproven because it is poorly defined. What is "transparency"? How can it be objectively measured?

(tried to post this a couple of times without succ...

2009-06-01T12:54:33.968-04:00

(tried to post this a couple of times without success - apologies if it gets posted more than once)

Nate, I know you didn't construct the table, but there's something wrong for "republican" and "None." As presented for republican, z=coef/se(coef)=4.2 (not 0.42 as is indicated in the table), which would be significant at the 5% error level. Moreover, exp(coef)=1.6 (not 1.05) which would indicate a much larger effect. I'm assuming it's a keying error on the coefficient since changing .47... to .047... makes everything else match. The same issue is true for "None."

I wish people (especially Nate, who knows better) ...

2009-06-01T12:41:01.801-04:00

I wish people (especially Nate, who knows better) would be more careful when describing p-values.

The p-value is NOT "the probability that the outcome occured due to pure randomness." ALL outcomes of statistical tests are entirely due to randomness, by definition, since they are the result of random processes.

If the p-value is 0.05, it is NOT the case that there is a 95 percent chance that the hypothesis is true. Only Bayesian theory can give such a probability, since only Bayesians are allowed to put probabilites on hypotheses. Frequentists are not allowed to put probabilities on hypotheses, and p-values of the sort used in this article are a frequentist construct.

The only thing you can say about a p-value is that if you observe a particular p-value, say 0.05 for definiteness, then in a large set of similar experiments, assuming that the null hypothesis is TRUE, only 5% of the tests would result in a p-value that was as small as or smaller than 0.05; that is, 0.05 is the Type I error rate (horrible terminology, but we are stuck with it) for a series of experiments that may or may not have been performed where the null hypothesis is actually true.

I once saw a fantastic signature line that directo...

2009-06-01T11:17:27.179-04:00

I once saw a fantastic signature line that directorblue should read, absorb, and use as a guide. "The plural of anecdote is not data.

Oh, yeah. I heard that there's a Chrysler dealership in Hawaii that's owned by a guy whose mother was a nurse at the hospital where Obama was supposedly born, and he got to keep his dealership because she knows the truth about Obama's birth certificate. (That should get repeated on Rush in about 10..., 9..., 8...)

directorblue - Thank you for responding to logic...

2009-06-01T11:07:53.426-04:00

directorblue -

Thank you for responding to logical rebuttal of your claims by merely repeating your dismissed claims, unmodified.

Others -

Even if there were some rational statistical analysis that showed that Republican-donating car dealers were more likely to be closed than Democrat-donating car dealers, which there is not, even that would not indicate a conspiracy. A third factor could be the driver.

Dealers would be likely to support parties whose legislation might be expected to be congruent with their business strategy. Thus, even among those who deal mainly in Chrysler cars, the subset who favor large gas-guzzlers might, for example, be both more likely to donate to Republicans, and more likely to have suffered poor sales of late and thus be closed.

However, again, there is no evidence to suggest excess closure of Republican dealerships.

As Nate mentioned, car dealers as a total group are and would be expected to be, overwhelmingly Republican donors.

We also need to remember that many business Republican donors also make small donations to a Democrat.

It's a moot point since the regression analysis depicted produces no significant results anyway, but one would need to control for the fact that "Clinton donor" car dealers might also be heavier donors to McCain.

Bravo, Nate. Posts like these are why I come to t...

2009-06-01T11:03:11.557-04:00

Bravo, Nate. Posts like these are why I come to this site.

that's "Dealships with crappy sales were more like...

2009-06-01T10:45:04.866-04:00

that's "Dealships with crappy sales were more likely to be closed"

I'm sure this has been pointed out, but has any te...

2009-06-01T10:44:22.200-04:00

I'm sure this has been pointed out, but has any tested the hypothesis:

"Dealships with crappy sales were mire likely to be closed"

or

"Dealerships close to other dealerships were more likely to be closed"

If either of those turns out to be significant, we can tell the magical mystery hypothesis to shove it.

Interesting comments; the site really attracts a w...

2009-06-01T09:25:39.268-04:00

Interesting comments; the site really attracts a wonderful collection of sentients.

As I think at least one person said, it's easy enough to look for one significant result on a post-hoc basis; you just use a different test and/or a different measure of significance. I can eyeball those data and know that result is purely random.

You do, however, write off "data dredging" too glibly. As per the roulette example, it is probably the best tool we have for generating interesting hypothesis from messy data. You just have to follow it up with a separate, independent test. Too bad that's wicked expensive.