So you killed your kid while cleaning your gun.

[Kable]

@Trinopus

Why? He told the truth, and you’re still lying.

OK, you are still too stupid to recognize truth. Lots of people are like that.

[Kable]

@Measure_for_Measure

Huh. Well, then I’m disappointed in Kable. Somebody who has taken 1 semester of college statistics can be forgiven for not appreciating multivariate analysis, but they really should know about confidence intervals.

Tell us what the most likely value is of a confidence interval? Is it the zero effect value off to one side? Is that value 95% likely? Cause that’s what Hentor said.

[Damuri_Ajashi]

@Euphonious_Polemic

If you really don’t understand what numbers and mathematics say, then it’s quite hard to have a logical conversation with you.

I think you should read the thread before you go any further.

@Trinopus

That makes sense intuitively…but not mathematically. It “feels” like it ought to be true. But as noted above, math doesn’t allow “eyeballing.”

What you’re looking at is a kind of meta-statistics, where one could (entirely validly) explore the distribution of data within the confidence interval. This, however, would require taking a new poll of the data, because, as presently given, we don’t know that. The data might all be clumped tightly against one end of the interval…or clumped tightly against the other end.

Mathematics doesn’t deal in “it might be.”

I asked about the normal distribution earlier on in the thread and I’m not sure if I was answered. If this is a meta analysis, then sure, I can see how the distribution would be skewed. Is that what happened? Does Kellerman provide us with the raw data to check on this or are we asked to take his word for it?

@Measure_for_Measure

Good catch: multivariate analysis is a pretty important concept actually, as it allows social scientists to do something analogous to a controlled experiment.

I thought that was what multivariate analysis was but, I don’t understond how this is accomplished (I suppose I don’t really need to understand as long as everyone who does understand agree on the results).

Another good catch. It might be better to use the phrase, “Statistically discernible”, but unfortunately the proper technical term is “Statistically significant”. A statistically significant effect can be miniscule in size-- all significance means is that the investigator rules out that what he is measuring was produced by chance (or if it was produced by chance, that would only happen 5% of the time or whatever ). Conversely, a large effect can be statistically insignificant if the power of your test is weak. A herd of elephants can have a pretty devastating impact on a one acre farm, but if the viewer is on the moon, it won’t be clear to him what is happening. The magnitude of the effect is high, but crude measurement prevents it from being confirmed.

There’s probably some Bayesian argument that a gun advocate could make, but it wouldn’t involve challenging the integrity of the author of the original piece.

Well, I know all this stuff is a tangent but I appreciate the comment. I work with actuaries from time to time but they might as well be a black box.

[Measure_for_Measure]

@Kable

Tell us what the most likely value is of a confidence interval? Is it the zero effect value off to one side? Is that value 95% likely? Cause that’s what Hentor said.

1. It appeared to me that you didn’t grok to the notion that if the confidence interval contains the zero value (or unity in the case of an odds ratio) then the phenomenon in question isn’t statistically significant.

2. To answer your question, it depends upon the shape of the underlying distribution. Usually they are taken to be symmetric, but not always.

3. More to the point, it seemed to me that you were pouncing on a technical imprecision more than anything else.

@Damuri_Ajashi

I thought that was what multivariate analysis was but, I don’t understand how this is accomplished (I suppose I don’t really need to understand as long as everyone who does understand agree on the results).

Haven’t read the paper. Presumably a variant of an ordinary least squares regression is used. Which is a black box if you haven’t taken 2nd semester statistics and frankly is typically recalled as such even if you have. (For those who have taken 2nd semester statistics and more, I’d recommend googling “Regression Anatomy”: there’s a new Stata command which presents a nice graphical presentation of models with more than one explanatory variable. This geek was thoroughly enthused.)

[Kable]

@Measure_for_Measure

1. It appeared to me that you didn’t grok to the notion that if the confidence interval contains the zero value (or unity in the case of an odds ratio) then the phenomenon in question isn’t statistically significant.

No, I gather the concept of statistical significance just fine.

2. To answer your question, it depends upon the shape of the underlying distribution. Usually they are taken to be symmetric, but not always.

That’s what I figured. So for example when Kellermann found the crude odd ratio for homicides with a shotgun in the home to be 0.7 CI (0.5-1.1) the difference was not statistically significan but still the most likely value is something like 0.7, and certainly there is not a 95% chance that value is 1.0 like Hentor said.

3. More to the point, it seemed to me that you were pouncing on a technical imprecision more than anything else.

“Technical imprecision” is a nice euphemism for “bullshit.”

[mayaonfiya222]

This pisses me off. Even when I’m watching a movie with an ‘unloaded gun’, I’m always thinking, STOP POINTING THAT THING AT PEOPLE. It’s common sense. If you’re going to clean your gun or do anything with your gun, do it in private. What the hell is wrong with people? Even at my stupidest, I can’t imagine doing something like this. That being said, that father will have to live with what he did…and although it was his fault, it’s still pretty sad.

[Hentor_the_Barbarian]

@Measure_for_Measure

… there’s a new Stata command which presents a nice graphical presentation of models with more than one explanatory variable. This geek was thoroughly enthused.)

I would be super-pumped about it, but I’ve found that most of the exciting and time-saving post-estimation stuff Stata announces, I end up being unable to use. Most of what I do involves panel data with count variables (and thus my modeling uses GLM with clustering or XT commands) and that seems to cause problems for all the cool commands. But Stata is my world, and I am ever the optimist, so I will look for this.

Basically I just need to find an area with nice normally distributed variables of interest instead of my current niche.

On the point at hand, I suppose the answer is “just another internet jackass.” In a nutshell, his position is that the authors of the paper are frauds pushing a biased agenda, that the design is flawed and the data is suspect, that the whole thing should be dismissed EXCEPT for the non-significant univariate association between two particular variables and the outcome.

[Cheshire_Human]

@Measure_for_Measure

… I’d recommend googling “Regression Anatomy”: there’s a new Stata command which presents a nice graphical presentation of models with more than one explanatory variable. This geek was thoroughly enthused.)

[hijack] Hmm… Thanks! Is “Stata” OpenSource/Freeware? Can I get a copy without paying? I could use that, but not enough to be worth paying for. These days I only do stats on hobby-type shit I’m not willing to sink any real money into. PM me. Don’t want to hijack a good Pit thread. [/hijack]

Ok, everyone. Back to ragging on Kable (who doesn’t know ‘stat’).

[Cheshire_Human]

@mayaonfiya222

This pisses me off. Even when I’m watching a movie with an ‘unloaded gun’, I’m always thinking, STOP POINTING THAT THING AT PEOPLE. It’s common sense.

This pissed me off when I was at the Mall in DC a few years ago, seeing the (then) new Korean War Memorial. One of the statues is hoofing his way through mud, holding his rifle in one hand, with his finger on the trigger. The first thought that went through my head was “What is your major malfunction, soldier?” At least he wasn’t pointing it anywhere someone might get hit. He’s pointing it somewhere in the mud about a meter or so to his front-right. My mother told me to lighten up. My brother, who’s been through Army basic, too, agreed with me. And now it sits on the National Mall. Really Bad weapon-handling, by US soldiers, going back more than my lifetime. And it’s now immortalized in a nationally supported statue. Shit.

[Trinopus]

Remember the really dumb scene in Die Hard II (yeah, yeah, dumb movie) where Bruce Willis’ character is in a room full of armed guys, and demonstrates that the gun he has only fires blanks…by firing it at a guy? And everybody just stands there and gapes in silent astonishment…

Our movies are really, really rotten sources of gun education…

[Measure_for_Measure]

@Hentor_the_Barbarian

I would be super-pumped about it, but I’ve found that most of the exciting and time-saving post-estimation stuff Stata announces, I end up being unable to use. Most of what I do involves panel data with count variables (and thus my modeling uses GLM with clustering or XT commands) and that seems to cause problems for all the cool commands. But Stata is my world, and I am ever the optimist, so I will look for this.

Basically I just need to find an area with nice normally distributed variables of interest instead of my current niche.

Well, the highlight of the article IMHO was conceptual: they were building off of the book Mostly Harmless Econometrics: An Empiricist’s Companion, which I’ve owned for a while but have not even opened. :smack:

Anyway, here’s one passage:

When the model includes more than a single independent variable, there is no straightforward equivalent for the estimation ofß, and the same bivariate scatterplot between the dependent variable and the independent variable of interest becomes po-tentially misleading because, in the general case, the independent variables are not orthogonal between them. Consequently, most econometric textbooks limit themselves to providing the formula for the ß vector of the type

Beta = (X’X)[sup]-1[/sup]X’y

and drop altogether any graphical depiction of the relation of interest. Although com-pact and easy to remember, this formulation is a sort of black box because it hardly reveals anything about what really happens during the estimation of a multivariateOLS model. Furthermore, the link between theßand the moments of the data distribution disappears, buried in the intricacies of matrix algebra.

Luckily, an enlightening interpretation of the ß’s in the multivarate case exists and has relevant interpreting power. It was originally formulated more than 70 years ago by Frisch and Waugh(1933), revived by Lovell (1963), and implemented in applied econo-metrics by Angrist and Pischke(2009) under the catchy phrase “regression anatomy”. According to this result, given a model withKindependent variables, the coefficient ß for the kth variable can be written as

[Magic equation]

The result is striking because it establishes the possibility of breaking a multivariate model withKindependent variables into Ksimpler bivariate models and also sheds light on the machinery of multivariate OLS. This property of OLSdoes not depend on the underlying data-generating process or on its causal interpretation: it is a purely numerical property of the estimator that holds because of the algebra behind it.

Ok, I’ll transcribe the magic equation here:

Beta(k) = Cov(y(i),x~(k,i)) / Var(x~(k,i))

Ok, let me put it another way. Say we’re running Y against a 3 variables: xA, xB and xC. Regress xA on the other two x’s and take the residual. If I understand this correctly if you run that residual against y (without the other x’s and I assume without a constant) that will give you the multivariate estimate for Beta(xA). Somehow I had missed that trick in my studies. Now I might have mangled something here - but the point is that you can turn a Y vs k indep variable regression into a set of k meaningful scatterplot charts. Admittedly they are not simple to read, but they aren’t overwhelmingly difficult either.

I trust there’s a kludge for other regression techniques. I have no idea whether they would be valid. (Though it would be straightforward to do a quick test.)

I’ve probably screwed something up in this presentation. But I found this paper pretty intriguing. It was buried in the current Stata Journal (2013) 13, Number 1, pp. 92–106. I think it deserves some attention.

@Cheshire_Human

[hijack] Hmm… Thanks! Is “Stata” OpenSource/Freeware? Can I get a copy without paying? I could use that, but not enough to be worth paying for. These days I only do stats on hobby-type shit I’m not willing to sink any real money into. [/hijack]

No, not legally. I suspect R might have something soon though.

Stata

R (programming language)

I googled this link, but did not sign in:

What are the respective advantages of Stata, R, MATLAB, and other packages...

ETA: Cite! Regression anatomy, revealed
by Valerio Filoso

[Cheshire_Human]

@Measure_for_Measure

… I suspect R might have something soon though.
Stata - Wikipedia
R (programming language) - Wikipedia

I googled this link, but did not sign in:
http://www.quora.com/Statistics-Software/What-are-the-respective-advantages-of-Stata-R-MATLAB-and-other-packages-for-performing-statistical-analysis

Thanks! I think I might still have a copy of R for OS/2 that I never looked very closely at (got it free, and didn’t have any use for it at the time). I’ll have to see if I can dig it out.

[Hentor_the_Barbarian]

@Cheshire_Human

Thanks! I think I might still have a copy of R for OS/2 that I never looked very closely at (got it free, and didn’t have any use for it at the time). I’ll have to see if I can dig it out.

I love the idea of getting these programs just to play with them, but do you have any data? Without data, they’re kind of useless.

Sure, you can generally find dummy data sets, but driving them around doesn’t seem like much fun.

[Cheshire_Human]

@Hentor_the_Barbarian

… do you have any data? Without data, they’re kind of useless.

Several 10s of thousands of horses past performances. Hundreds of thousands of races. I think I have some data I can play with.

[Measure_for_Measure]

@Cheshire_Human

Thanks! I think I might still have a copy of R for OS/2 that I never looked very closely at (got it free, and didn’t have any use for it at the time). I’ll have to see if I can dig it out.

May as well download the most recent version: http://www.r-project.org/

@Hentor_the_Barbarian

I love the idea of getting these programs just to play with them, but do you have any data? Without data, they’re kind of useless.

Sure, you can generally find dummy data sets, but driving them around doesn’t seem like much fun.

Not sure if that was a pun, but auto.dta isn’t bad.

@Cheshire_Human

Several 10s of thousands of horses past performances. Hundreds of thousands of races. I think I have some data I can play with.

That sounds awesome.