Via William Sjostrom I find that Mike Adams has distilled Lancet denialism down to its essence. He quotes from the New Scientist's description of the Lancet study:
"The invasion of Iraq in March 2003 by coalition forces has lead to the death of at least 100,000 civilians, reveals the first scientific study...of almost 1000 households scattered across Iraq."You read that correctly. A scientific study "of almost 1000 households" determined that we killed 100,000 civilians in Iraq before the November election.
Believe it or not, Adams is a criminology professor and he can't conceive how a survey could estimate the number of excess deaths.
More like this
Note for visitors from Daily Kos: 120,000 is an estimate of the
number of violent deaths. The total number of extra deaths as a result of the war is very roughly
200,000 once you include the increase in disease and accidents since
the invasion. This number is more likely to be too low than too…
As my readers know, the reason why the Lancet study and the ILCS give different numbers for deaths in Iraq is because the studies measured different things over a different time periods. Of course, that fact isn't going to stop pro-war columnists from claiming that the ILCS refutes the Lancet…
Welcome to the 2004 Deltoid awards. Today we are giving out the Golden Rake Award, named in honour of Sideshow Bob and the rakes in the Simpsons Cape Feare episode:
How many other series would waste valuable prime-time real estate by showing a man whacking himself in the face…
In an earlier post on the IBC I wrote:
Sloboda says:
We've always said our work is an undercount, you can't possibly expect that a media-based analysis will get all the deaths. Our best estimate is that we've got about half the deaths that are out there.
OK, then why does the IBC page say "Iraq…
This must get boring, no? For one thing, "At least" and "civilian" remain distortions. "Nearly 1,000 households" alludes [again, appropriately] to the low power of the survey, the final resting point for all conversations on this study.
Most of the anti-Lancet study articles have been exercises in hackery. But one question bears repeating: why haven't there been more images of hospitals full of wounded civilian Iraqis? If there was an increase of 98,000 violent deaths, doesn't it stand to reason that there should be an equivalent increase in non-fatal injuries? If so, where are these people?
JR,
First, remember the survey shows an increase of 98,000 deaths, not all of them violent by any means.
Second, I'm not sure how many pictures of hospitals "full of wounded" you'd expect to see. I can think of a host problems with the expectation: The victims are scattered geographically and temporally; not all victims of violence will go to a hospital; US media doesn't seem eager to seek out or broadcast such photos; reporters are, by most accounts, pretty limited in where they will travel; in at least one case, the coalition has taken over hospitals to avoid pictures and reports from reaching the media; and so on.
Taking all this into account, how many such pictures would you expect?
See Riverbend for an incident just last month not covered by the Western media. Page down about nine paragraphs to see an example of how tragically overcrowded the hospitals in Iraq are.
Sorry, no pictures.
I'd like to add to Grogzwig's list another reason why we're not seeing pretty pictures from Iraqi hospitals: the US has a policy of attacking hospitals.
This is exactly what happened at the beginning of the campaign to liquidate Fallujah.
In another incident, the US ordered Spanish troops to attack a hospital with full force in order to root out insurgents. The Spanish balked and instead secured the hospital with special ops.
JR - thats an interesting point, but surely the answer is to conduct a more thorough survey and identify a closer figure to the truth than that of the Lancet study; not just make up a load of reasons why its wrong and trumpet it ad infinitum as so very many people have done.
I also note with something approaching humour that the same Mike Adams who thinks that a survey that finds 100000 deaths from a sample of 1000 homes is untrustworthy is more than happy to totally back John Lott in his rather smaller (and yet larger extrapolations) surveys - indeed he wrote "Some of my colleagues disagree with his recommendations for public policy but are completely unable to specify any flaws in his research or in his logic. In fact, one of them once told me that he was very disturbed by data indicating that right-to-carry permits seemed to be reducing the amount of crime in jurisdictions that had made them available."
http://www.townhall.com/columnists/mikeadams/ma20040624.shtml
Well one reason there's been so little reporting from hospitals is that ever since, IIRC, late 2003 reporters have been required to obtain permission from the authorities to conduct interviews with medics.
This may have been prompted in part by reporting like this:
"why haven't there been more images of hospitals full of wounded civilian Iraqis?"
There used to be such images in abundance in the Arabic and some other non-U.S. media. This did not, of course, work well for the Bush administration (inflames the Ayrab masses to see all those wounded and dead Eyerackians, doncha know), so they took a variety of steps to make it difficult to obtain and/or broadcast these images, and to discourage the press and the media from attempting to do so.
Effective quoting JoT.
The 'no images of hospitals full of wounded' and 'but we use smart weapons' rebuttals pop up far too often for my liking. Utterly ridiculous heuristics/logical abductions. Laser guided or not, a 500 pound bomb still creates the same size crater - and doesn't first stop to ask civilians to get out of the way.
In this case I think you ought to look at the page by Jane Christensen that Mike Adams was really attacking, which was considerably more crazy than just about all the Lancet denialism Tim attacks. Statements like "More than 100,000 Iraqi Civilians Killed by US" are misleading, and while Mike may have gone too far in attacking the Lancet study itself rather than just the misrepresentations of it, given this sort of distortions one can understand if people get confused.
The "more than 100,000 civilians" comes from the New Scientist article so we can't blame Christensen for that bit of imprecision. She also obviously joking when she says she is the one on the left in the picture of some terrorists. (Adams doesn't seem to realize this.) She also links the the Republican National Committee, so you can't say that she agrees with everything she links to, which seems to be Adams' main argument.
A report on the state of several hospitals in Iraq 2/2005
http://www.medpeace.org/news/mfp-report-civilian-health-in-iraq.htm
Things are not good.
"She also obviously joking ..."
Why do you say that?
The midrange figure for violent deaths in the Lancet study is 60,000. Some of the critics of the study that I've seen around here seem to acknowledge that this is realistic figure and they've only tried to claim that most of them are either insurgents or the civilian victims of insurgents. So you'd have the same "where are the wounded?" question in that case.
It's not uncommon to have huge uncertainties in the death toll in a war. Try pinning down the number of civilians who died in Vietnam--there are plenty of widely varying figures around (from a few hundred thousand to over 2 million), but it's hard to tell if any of them have real facts behind them. I do remember reading that in Vietnam, one doctor knew of more cases of plague in his province than the South Vietnamese government claimed for the entire country. So governments tend to lie sometimes. What a shock.
Alistair Horne's book on the French/Algerian conflict has a page on the death toll there and it's uncertain by a factor of 3, with most of the uncertainty being the number of civilians. The accounting he gives is revealing--the French seem to have kept count of the number of civilians murdered by the Algerian guerillas, but there are only impressionistic estimates of the much larger number of civilians killed "accidentally" (Horne's word) by the French military.
It is possible that Jane Christensen's page is a parody, but as such it is so good that it is impossible to separate from that of a real crackpot. As such I can't really fault Mike Adams for taking it seriously.
we can't blame Christensen for that bit of imprecision.
Who is accusing Christensen of anything? You quoted Adams quoting the New Scientist. Adams is ridiculing the conclusion as presented, and he is correct to do so since "at least 100,000 civilians" is not the conclusion of the study. This is a glaring factual error, not a "bit of imprecision."
This post is entitled "more Lancet denial." Where in the article does the author deny any of the much-discussed "core findings" of the study? I see only a glancing reference to a flagrant misrepresentation of its conclusions. Miraculously, you find fault with Adams and not with the New Scientist or Christensen for uncritically parroting this falsehood.
If this is the essence of Lancet denialism (continuing to deny that the study has been misinterpreted by 99.9% of the known universe, or claiming that it doesn't matter) I think you may have gotten it right.
Pascal, the vast majority of the deaths were of civilians and the "at least" refers to the exclusion of Falluja. I don't think it is a good way to describe the findings but it is not a "glaring factual error". Adams denies that a survey of 1000 households could find such a thing; he does not correct the misstatement of the findings.
JF Beck, it's an obvious joke. She has a picture of terrorists on her page. "Which one is you?" "The one on the left." If you want proof, that picture comes up as the second result of a Google image search for "terrorist".
Shirin and dk.au have hit the nail on the head. The reason we aren't seeing masses of injured or dead Iraqi's is because our wonderful corporate media does not want us to see them. It's bad for our image, you see.... For example, the Americans did not escort the tens of thousands of civilians who lived in Falluja out of the city prior to flattening the city. It literally bombed them out. Yet how many images of these people have we seen on TV or in the mainstream corporate press in the west? These refugees must be somewhere (those 2,000-3,000 at least that the Americans didn't kill, that is). Perhaps up to 100,000 of them, whose homes were levelled, are probably living in tents in the desert. But I have yet to see these harrowing images.
The Americans also created some 300,000 refugees in Afghanistan who huddled in the Kazakh (hopefully my spelling is correct) camp. But our exalted media said virtually nothing about them. How many times does it have to be said: that the Iraqi people mean nothing to the western establishment? Just like the Afghans, Nicaraguans, Indonesians, Guatemalans, Viet Namese, and people from dozens of other nations before them.
the vast majority of the deaths were of civilians and the "at least" refers to the exclusion of Falluja.
Neither portion of this statement is supported by any detail of the study, as you should know by now after all these threads. "At least 100,000" makes no sense with or without Fallujah and Fallujah, for the thousandth time, was excluded from the study and its conclusions. Furthermore, again, you have no idea what the study's confidence interval including Fallujah might be (its lower bound is almost certainly well below "zero excess deaths.") Civilians are not differentiated in the study, and it DOES make a difference whether the total there is 30,000 or 60,000 or 100,000. Have you understood any of the criticisms voiced on this page? Why do you continue to distort the findings of the study?
For his part Adams' criticism is confined to a single insubstantial remark.
Pascal, I do not and have not distorted the findings of the study. If you had bothered to read the study you would have found that the very first sentence of the findings states:
That is what you get if you include Falluja, and no, the CI does not include 1.
I already pointed out to you that civilians make up the vast majority of the deaths. Apparently you do not understand this.
By the way, could I make a sort of appeal to the denialist community?
Could we just take it as understood that Tim, me, Chris Lightfoot and the rest have actually read the Lancet study, and that therefore there is no point in trying to fool us about what it says? Please?
That is what you get if you include Falluja, and no, the CI does not include 1.
Perhaps you are still unaware that the "excess death" figure carries a confidence interval distinct from the "risk of death" confidence interval, and that the two CIs are calculated in completely different ways, as the 8 page summary (which I have also read) makes quite clear.
Dsquared, why are you not chiding Tim for [yet again] misrepresenting the conclusions of the study? Please explain to Tim what the inclusion of the Fallujah outlier is likely to do to the lower bound of the "excess death" confidence interval as it seems not to be sinking in. 100,000 is simply not the minimum figure represented by the study. Grafting the unqualified (and excluded) Fallujah cluster to the midpoint of the ex-Fallujah clusters in order to justify such language is conceptually incoherent.
there is no point in trying to fool us about what it says?
Right, whereas when Tim claims:
* that the study's represented minimum is 100,000
* that the excess death CI is algebraically related to the risk of death CI, and that were it to include fallujah it would still exceed zero
* that "the vast majority" of the 100,000 midpoint estimate are civilians
* That the "New Scientist" headline is "a bit of imprecision"
who do you imagine is [still] being fooled? It isn't "denialism" to draw attention to yet another misreading of the study's results! It IS denialism to defend such misrepresentations and offer up falsehoods in their defense.
Tim, perhaps you need to take a break from this Lancet study. You are becoming so zealous in finding any heresy that you sometimes become as unreliable as the people you attack. The statement "The invasion of Iraq in March 2003 by coalition forces has lead to the death of at least 100,000 civilians" is wrong, and your defence is silly. You'd be the first to attack anyone who used similar sloppy arguments to discredit the study. The text of the New Scientist story explains some of the details, but the headline is wrong so why not just admit it?
Re the lack of documented wounded in Iraqi hospitals: Without doubt, part of the reason for so high a mortality rate is the post-war lack of functional hospitals in Iraq which could have made a huge difference for both illness and trauma victims.
Thomas Palm (to Tim):
Tim (only a few lines above):
Ditto to Ombudsman.
However, the statement "The invasion of Iraq in March 2003 by coalition forces has lead to the death of at least 100,000 civilians" has been attributed (in a separate Deltoid comment thread) to the Lancet authors from a letter they supposedly wrote. I can't see the grounds for that statement, if the attribution is truly accurate. It's the only significant criticism I've seen so far of the Lancet authors, and it's not in the report itself.
Ombudsman, Tim is just a guy with a Lott of hobbies.
Pascal,
If the risk of death in Iraq as a whole actually fell, or if it increased by any factor less than 1.6, the probability of obtaining a sample (inclusive of Fallujah) such as the one we have is 2.5%. Assuming that one-in-forty chance has not transpired, Iraqi mortality has risen by at least 60%. That in turn means that the number of excess deaths is very large and certainly greater than zero.
However it is conceivable that the program used by Roberts et al. to generate the ex-Fallujah confidence interval for excess deaths would give an anomolous lower value (maybe even a negative value) if the Fallujah figures were fed into it. That seems quite unlikely to me. I will believe it when I see an argument based on distribution theory.
But even if that is the case, so what? It is obvious that true mortality doesn't fall when battlefields are included in the total, whatever the computer may say.
Assuming that one-in-forty chance has not transpired, Iraqi mortality has risen by at least 60%.
Fallujah was excluded from the study, so the 1.6x figure is not part of the study's conclusions. Why is this so hard to get across? Fallujah is not part of the study results! Believe me, it's better this way, because including it nullifies the "excess deaths above zero" argument altogether.
That in turn means that the number of excess deaths is very large and certainly greater than zero.
As I said and as the study summary makes clear, the excess death and risk of death CIs are calculated according to different methods. "Certainly greater than zero" is not accurate, at least not according to the conclusions of the study. The study does not include Fallujah. If it did, the CI for excess deaths would (I am convinced on the basis of examining many similar data sets) encompass zero. You can try getting such a CI out of the authors (I did.) I doubt you'll succeed.
This is why the presentation of the data is highly political and deceptive: a CI and mean figure are presented for risk of death , but not for excess death.
See what introducing a sample 66 times the mean will do to a distribution's standard deviation, and you will see that a sub-zero excess death number is more than plausible (we'll never know for sure without knowing exactly how such a figure was normalized/bootstrapped.) An outlier 66 times the mean result of a set of 33 can blow the standard deviation out tenfold if not more.
Sneaking Fallujah into the discussion does not absolve Tim or the authors of the obligation to state either that cluster's unique CI, or the study-wide CI inclusive of the cluster. If you intend to cite CIs and quantitative measures of likelihood, you must do so consistently, or not at all. Confining your discussion to the unaugmented risk of death, or the excess deaths ex-Fallujah is honest. Claiming a minimum of 100,000 deaths is dishonest and unscientific, let alone 100k "civilian deaths" or anything similar.
How about if I make one last go at playing whack-a-troll?
First, the Fallujah cluster was not, I repeat, NOT excluded from the study nor its conclusions, no matter how many thousands of times one says otherwise. The Fallujah cluster was only excluded from calculating the extrapolated expected value (and associated CI) of excess deaths, which is appropriate. Whoever insists otherwise really does not know how to read a research paper.
Second, the Fallujah cluster was excluded from calculating the expected value of excess deaths because it was an outlier. However, it was not an outlier due to some kind of measurement error or other form of false reading. It was recognized that it was an outlier because it was measuring a phenomenom different from the *baseline* excess deaths that the rest of the clusters measure, a phenomenom that while important to note in calculating TOTAL excess deaths, had to be distinguished from the baseline that the rest of the clusters measured. In short, the authors of the study recognized that Fallujah was a battle field, and so handled this cluster differently.
Finally, I'm glad that some denialists have finally managed to slog their way through the Lancet paper. However, just as those reading Shakespeare for the first time are unlikely to comprehend what is transpiring, those reading peer reviewed research for the first time are unlikely to grasp important nuances, like, eg, statistical analysis.
Not that I expect any of this to sink in or be understood by those for whom the Lancet results are a threat to their political agendae.
unlikely to grasp important nuances, like, eg, statistical analysis.
Man oh man you guys just can't get away from pompous appeals to authority, and this one's as hollow as they come. The conclusions of the study a.) don't include a "baseline" distinct from the lower bound of the 95% CI b.)don't include the Falluja cluster, else they would have declared an estimate of 300,000 excess deaths with a CI around 1 million deaths wide centered upon this number.
However, it was not an outlier due to some kind of measurement error or other form of false reading.
I said nothing about a "false reading." This makes me think you don't understand the character of my objection at all. t I'm interested to hear why you think it credible that 200,000 Iraqis might have been killed in Anbar province alone, seeing as that amounts to over 1/4 its entire population.
slog their way through
I have read that paper perhaps 10 times through (it's only 8 pages so not much of a slog.) Not that it matters one bit, but I read a lot more complicated stuff every day, some of it involving far more difficult mathematics and stochastic analysis than anything contained [or merely alluded to] here. You seem not to have read it once, or you would understand that the Fallujah cluster is not included in the study's reported findings.
The cluster was excluded from the reported results for one reason, and it had nothing to do with the method of sampling in Fallujah or that it was a "war zone" (come again?) Once again, Les Roberts: "not giving a number or a range that could have been discredited and undermined the validity of the study seemed then and now like better service to the Iraqis." Kind of says it all for me.
I continue to be amazed by the brazen way Pascal misrepresents the findings of the study. Pascal claims:
Actual findings, first sentence, already posted in this thread:
That is including Falluja. And he claims to have read the study....
Wow Tim, I didn't notice that passage before, thanks. Since the findings of the study clearly show that 300,000 excess deaths have occurred, we should certainly be pursuing a war crimes charge against Tony Blair and George Bush. I wonder why the press is running with that 100k number. Asleep at the switch, eh?
Or maybe the study found the 200k fallujah deaths occurred only for the purposes of the risk measure but not for the purpose of the excess death measure (except when they did, and then please don't ask about the confidence interval as you're probably not clever enough to understand how it was derived.)
Crystal clear, unimpeachable science.
Pascal,
300,000, or even 100,000 deaths are not necessary to bring war crimes charges against the perpetrators of this war. For starters they committed the the most grievous war crime of all when they launched a war of aggression. The destruction of Falluja was a war crime regardless of how many died, and might rise to the level of a crime against humanity. The use of cluster munitions in cities was a war crime. Cutting off the water and electricity to Basra, Falluja, and other cities was a war crime. Their recommendation and approval of the abuse and torture of prisoners is a war crime. And those are only the ones that come to mind immediately.
More war crimes:
Attacking hospitals and clinics, including bombing and destroying; taking over hospitals for use as bases for military operations, interrupting medical personnel while they are treating patients, physically abusing patients and medical personnel, dragging patients out of their hospitial beds, forcing doctors to abandon patients in the middle of surgery or birth, evicting patients and medical personnel from hospitals; attacking ambulances, medical and rescue personnel; denying rescuers and medical personnel access to the wounded; depriving children of education by taking over schools for use as bases.
Attacking refugees attempting to flee cities under U.S. attack; forcing people who are fleeing cities under U.S. attack, or threatened with U.S. attack to reenter the city; taking hostage family members in an attempt to induce surrender of wanted individuals; imposing collective punishment on families, neighborhoods, or entire communities for the acts of one or a few members.
Scrapping the country's constitution and imposing a constitution composed by the occupying power; scrapping the laws of the country and imposing laws devised by the occupying power for the benefit of the occupying power; imposing a new and completely different economic system (dubbed "economic shock treatment"); dismissing hundreds of thousands of government employees; imposing its choice of a form of government; dictating changes in the education system and curriculum; changing national symbols (e.g. the flag).
Wow Tim, I didn't notice that passage before, thanks
WHAT?!?!
It is the first sentence of the "Findings" section! It is on the eighth line of the front page! It is the first sentence reporting actual results (as opposed to "Background" and "Methodology") in the whole paper! It is, as far as I can see, the most prominent sentence in the paper and certainly the first one that a casual reader's eye would fall upon! This is, quite literally, the equivalent of not having noticed that A Tale of Two Cities contains the sentence "It was the best of times, it was the worst of times"!
Do you not realise, Pascal, that your little tantrum over this issue has rather materially damaged your credibility? A bit more humility might be in order from here on in.
Other than that, your paragraph about "300,000" deaths represents you pretending not to understand issues which have been explained to you, patiently and at length, and your sarcasm about confidence intervals is misplaced; it's quite clear to me that someone who can't read a "Findings" section properly isn't "clever enough" to understand a bootstrapped confidence interval, even though I explained to you in layman's terms.
I'll no longer be responding to Pascal (although I may occasionally post a link to this comment so that people can see why I'm following this policy). Other commenters can do what they like.
Pascal,
Thank you, that's the funniest thing I've seen in ages. Have you considered comedy as a career? You could do a series based on life in a marketing consultancy, call it Fawlty Research or something like that.
Wow, some people really know how to read things correctly! And wow, how they want some of us to be wowed and others to be cowed by it.
To which one can only respond: Wow! Wow! WOW!Some may go so far as to bark Bow-wow! Bow-wow! Bow-wow! to proclaim their place in the clever clog's club, but that's not for me.To me it looks like the time has come to give thanks where due, and bow out.
The authors DO throw out the Fallujah cluster data, as an outlier, when reporting their calculated extrapolations based on the raw data.
Disputo, are you the author of this comment? Is a confidence interval for "risk of death" a calculated extrapolation based on raw data?
dsquared, the core of this argument has been the selective representation of Fallujah within the study summary. Up to this point I have generously interpreted the declared quantified conclusions of the study to exclude Fallujah, despite their discussion in the summary, and incorporation within the section labeled 'findings.' Obviously I read that passage many times, since I have referenced it again and again in this thread and earlier threads. Up to this point I had imagined that we all were in agreement that this cluster and its associated confidence intervals did not form a core finding of the study at all. I seemed not to be alone in believing this, note disputo's comment above:
The authors DO throw out the Fallujah cluster data, as an outlier, when reporting their calculated extrapolations based on the raw data.
"Risk of death" is a calculated extrapolation based on raw data. The confidence interval for "risk of death" is a calculated extrapolation based on raw data. dsquared now seems to have no problems admitting Fallujah to be part of the study conclusions in this context.
"better start mugging up on Feynman-Kac derivatives" -- anyone capable of writing this comment should not be lecturing anyone on humility.
After a bit of pompous fluff about your career as a derivatives mid-office quant, dsquared had this to say:
"The sensible thing to do...is to just give the baseline number and report the Fallujah cluster as a separate piece of information. Which is what they did."
dsquared: the Fallujah cluster is "a separate piece of information."
It may be useful to clarify some terms:
The relative risk ratio for the population as a whole is a parameter. An estimate of that parameter can be obtained from a sample. The formula or algorithm used to obtain it is an estimator. An estimate is not an extrapolation.
Extrapolation means the use of a parameter estimate to make an inference outside the appropriate range. For example, suppose researchers calculate the average height-to-weight ratio of a randomly-selected group of individuals, the shortest of whom measures 5' 6" and the tallest of whom measures 6' 0". Using that figure to guess the weight of somebody whose height is outside that range is extrapolation.
We now return you to our scheduled bickering.
Kevin Donahue:
If the risk of death in Iraq as a whole actually fell, or if it increased by any factor less than 1.6, the probability of obtaining a sample (inclusive of Fallujah) such as the one we have is 2.5%. Assuming that one-in-forty chance has not transpired, Iraqi mortality has risen by at least 60%. That in turn means that the number of excess deaths is very large and certainly greater than zero.
Is your "in turn" reference here to excess deaths (based on the mortality estimator an extrapolation ? Is it based on a core conclusion of the study? Does the language alluding to an estimate "probably more" than 100k employ similar (entirely fallacious and statistically incoherent) reasoning? Does it form any part of the conclusions of the study?
Mightn't a confidence interval around this estimator be fairly termed an "extrapolation?" Is the confidence interval addressing mortality a core conclusion of the study? I notice you seem happy to discuss the confidence interval related to the mortality rate, but not so interested in the CI related to excess deaths. Why not? Are both core findings of the study, or neither, or only the one?
dsquared is probably off pricing some "Feynman Kac derivatives" so I won't trouble him for a response any more. Since his MBA missed the section on how an outlier impacts overall volatility (or the confidence intervals of a heavily skewed distribution) maybe he has nothing to add. He doesn't seem interested in why it is appropriate to represent a CI for mortality and not for excess death in the conclusions of the study. Maybe he thinks the CI of a heavily skewed distribution is not so useful after all. Pity he's gone silent.
Here's a bit of easy math for the non-mid-office quants or junior faculty: the mortality estimate pre- and post Fallujah are 1.5 and 2.5 (an increase of 2/3.) The CI for this figure then more than doubled in width. What might the confidence interval for excess death do if the excess death estimate were tripled? Do you care to know the answer? Neither did R. Garfield and Les Roberts (we think.) "It would be very wide." It might "discredit the study."
Why are the Fallujah-inclusive risk and its CI part of the core findings of the study if the excess deaths and its CI are not?
Wow Tim, I didn't notice that passage before
Ombudsman, my comment was sarcasm. It was directed at the disingenuity and dishonesty of representing an extrapolated fallujah CI and extrapolated mortality rate (all rates obviously constituting extrapolations!) as part of the formal conclusions of the study.
The reason I am picking on dsquared the derivatives back/mid-office quant is because as a self-declared expert on time series analysis and Ito calculus and skewed distributions and volatility and like matters, dsquared should be in a good position to comment intelligently on statistical modeling and the abuses thereof, rather than attempting ineptly to use his [meager] credentials as a cudgel.
There comes a point in every discussion when you realize the person you are talking to isn't interesting in educating or informing, but instead making you feel worse. What you do at that point is left as an excersize for the reader.
interesting should read interested.
Pascal,
You must be a sort of Super-Kerryman. It is said that if you ask a Kerryman a question he answers with a question of his own. I ask you a question (or rather an attempt at clarification with a supplementary question tacked on) and you respond with no less than eight.
I wrote: "Assuming that one-in-forty chance has not transpired, Iraqi mortality has risen by at least 60%. That in turn means that the number of excess deaths is very large and certainly greater than zero."
Your 1st question: Is your "in turn" reference here to excess deaths (based on the mortality estimator) an extrapolation? .
No, it is a simple application of the principle that if you multiply a positive number by 1.6, the number you obtain is higher than the number you started with. 1.6 is the lower end of the Fallujah-inclusive CI for the relative-risk ratio. I am referring to the true number of excess deaths and using the assumption that the one-in-forty chance has not transpired, so that multiplication by at least 1.6 is justified.
Your 2nd question: Is it based on a core conclusion of the study? .
Of course; see above. But I am using words in their ordinary sense here. The study presents a finding for the rise in the risk of death (2.5 times) and a CI to go with it. If you have your own definition of "core conclusion" then maybe that isn't one for you.
Your 3rd question isn't a question at all. You suggest that I am engaged in "fallacious and statistically incoherent reasoning" and you invite me to agree that the Lancet authors are also. Don't give yourself airs. As to your 4th question: see answer to your 2nd.
Your 5th question: Mightn't a confidence interval around this estimator be fairly termed an "extrapolation?".
No. Statistics is not an easy subject where you can play fast and loose with definitions without getting into a muddle. An estimate is not the same as an estimator, a confidence interval is different from both and extrapolation is something else entirely.
With your 6th and 8th question you are back to your old obsession, core conclusions. See 2nd and 4th. There seems to be a pattern here.
Your 7th question: I notice you seem happy to discuss the confidence interval related to the mortality rate, but not so interested in the CI related to excess deaths. Why not?
I am interested in both. That is why I asked my question. You claimed that if the CI for excess deaths is calculated with Fallujah included it encompasses zero. That is an extraordinary claim, given what we know about the relative risk CI. You have provided no reasoning in support of your claim, only a reference to your experience of "examining many similar data sets." That is, to use your own words, "a bit of pompous fluff."
As for my supplementary question, I don't suppose you will ever get round to that.
I've explained the issue about confidence intervals a few times to Pascal and I'm not at this stage interested in giving a tutorial, but for any neutral readers the intuition is clear; if you get a bunch of observations between 200 and 300, then one observation of 5000, does that make you think that the true value might be 12?
does that make you think that the true value might be 12?
Appealing to intuition is exactly what you shouldn't be doing right now, BS artist. You should be trying to scrape together an argument that is coherent and scientific (good luck) that explains why a CI extrapolated from raw Fallujah data is included for risk of death and not for excess death, and why it constitutes a core finding of the study, and why no similar number was expressed for excess death, since as you well understand it would likely extend below zero. Glad to see we've pulled you out of retirement. I'm sure your explanation will delve lavishly into "standard epidemiological practice" which you know so well from googling in between coding IR option models that I would guess are super intuitive and badly inaccurate.
In fact, it is essential to the deceptive presentation of the study that your BS "intuitive argument" sound superficially appealing to the uninformed or uncurious.
But perhaps I give you too much credit. For the umpteenth time, amalgamating the means without the CIs is garbage. The CI is critically important information here, it is in fact your core [rational and cogent] argument, when you aren't sneaking peeks at Fallujah left and right and building it into what you declare to be the study's conclusions. I read and understood all your earlier crap about Fallujah coming from a different distribution altogether, thanks. It's part of MY point as well.
A confidence interval has been expressed [bootstrapped from outlier Fallujah data] that now supposedly is a core finding of the study. This is a major breakthrough, since now we deserve a full frontal view of the cluster under discussion, rather than just getting some suggestive peeks. A mean here, a confidence interval there = incoherent and deceptive garbage.
[for people unschooled as dsquared in SDEs and trinomial forests: Try 32 numbers between zero and three, averaging 1, with a standard dev of .5 or so. Then throw in a 66; the standard deviation jumps by a factor of 7. The 95% confidence interval now extends well below zero. That's the kind of observation we have here. This obvious and simple point has been made again and again, to quizzical stares from the crowd of supposed stats experts and their hangers-on, and demands for proof. I realize that the distribution suggested by the outlier does not lend itself to confidence intervals calculated under these assumptions: ANY confidence interval. I also have no idea what the bootstrap entailed here for either number and neither does dsquared. Which is why I asked Les Roberts and Richard Garfield, and they gave me the unsatisfactory answers noted above. I encourage someone to go get a straight answer. I will reward the first person to do so with a handwritten apology on nice stationery if I'm wrong.
Why does the CI for risk inclusive of the outlier stated as we have in the conclusions of the study (which bears NO functional relationship to the CI for excess deaths as the two were calculated under completely different bootstraps using different methods as the study makes clear) lend itself to such norm-based confidence measures? Clearly it has led to some misunderstanding among the less numerate among us (note Kevin's naive belief that the 1.6 CI for mortality risk connoted a non-zero lower boundary for excess death.) Note also dsquared's completely disingenuous appeal to "common sense" (which he as a master of stochastic calculus and all things Ito should know goes out the window in any stats-related discussion.)
The rest of Kevin's answer is confusion or obscurantism. Any "rate" entails extrapolation, especially a rate that may be used to extrapolate a minimum excess death number (which it can't, but you don't seem to care about that either.) "at least 1.6 ": WRONG. "fawlty research", "don't give yourself airs" - look who's talking, buddy. Try not fucking up a major factual point before saying stuff like this (you too, faker MBA.)
Doesn't simple kevin realize that dsquared's "2 distributions" argument is on a collision course with his and simple "in over his head" Tim's own "1.6 time sumthin give me the lower CI for excess deaths" blather? I hope dsquared's original stance prevails, and that d^2 himself snaps out of his own incoherence, cause I suspect he knows better, with all that expensive schooling behind him.
Pascal, you sound like a fricking angry baby. I guess it is natural to feel that way when everyone in the world is so wrong and you have to waste a million words to show 'em.
Try religion to see if a higher beign is more sympathetic and soothing. Or just tone it down.
The problem with Pascal's wager is that he postulates it likely that one could by chance have found a cluster to balance the Fallujah cluster but on the low side. Otherwise his argument about outliers flattening the distribution are garbage. He is assuming that Fallujah belongs to the distribution described by the other clusters.
If it is drawn from another distribution (the heavily bombed areas of Iraq) then 1. It should not be included in statistics describing that distribution and 2. The effect of including heavily bombed/damaged areas would be a bra curve not a flattened normal distribution.
So then the real questions are, given the distribution described by the study ex Fallujah, what is the probability that one would by chance have found a cluster with that number of excess deaths found in the Fallujah cluster. If that is very improbable, then you have to conclude that it does not belong, which is what Roberts et al said and that Pascal is peddling nonsense.
Pascal, if the relative risk of death is 1.6, the number of excess deaths is 0.6 times the death rate before times the population after. We don't know the exact death rate before so this doesn't, by itself, tell us the number of excess deaths. But the number zero has a curious property. Anything times zero is zero. So, if the relative risk is 1.0 the number of excess deaths is 0.0 times the death rate before times the population after i.e. zero. Since 1.6 is bigger than 1.0, the 95% CI of 1.6-4.2 including Falluja means we are more than 95% confident that the number of excess deaths is positive. I hope this helps.
The problem with Pascal's wager is that he postulates it likely that one could by chance have found a cluster to balance the Fallujah cluster but on the low side.
I said or implied nothing of the kind.
He is assuming that Fallujah belongs to the distribution described by the other clusters.
I said the exact opposite. I said that CIs applied to the cum-Falluja numbers are meaningless. Dsquared agrees with me here, right dsquared?
if the relative risk of death is 1.6, the number of excess deaths is 0.6 times the death rate
This is total baloney. How many times are you going to try peddling this lie Tim? The CI's bear no functional relationship to one another at all. Assume your notion of direct proportionality: what value of x would possibly make .1x = 8000 and 1.3x = 194,000. This is the second time you've tried making this claim, and in the intervening time it hasn't grown any more true. My epidemiology skills might be a bit rusty, but my algebra skills are just fine. So are my reading skills. "Loglinear" is still not "linear."
Pascal, your whole argument pre-supposes that the Fallujah cluster belonged to the same distribution as the other clusters. If it does not you are simply wrong.
Therefore you need to consider the probability of that cluster belonging to the distribution, and it looks to me that it lies somewhere between slim and none.
If it does not, your entire opus about how including Fallujah will flatten the distribution and drag the 95% CI through zero is blather.
Pascal asks:
Apparently you are unfamiliar with the advanced concept known as "rounding". Try x = 150,000. 8,000/150,000 = 0.1 (rounded to nearest 0.1) 194,000/150,000 = 1.3.
This is also missing the point. We don't know the precise value of the pre-war death rate -- it's one of the things the study measured, so the uncertainty in the number of excess deaths depends on the uncertainty in the pre-war death rate and the uncertainty in the relative risk. So you don't expect an exact linear relationship between the CIs. But as I explained before any number times zero is zero, so for the special case of 0 excess deaths and 1 relative risk, the pre-war death rate does not matter.
Eli, I don't know how many times I have to agree with you for you to recognize it. Real slow now: CIs applied to the cum-Falluja numbers are meaningless Understand? NO CIs incorporating outlier Fallujah data have any place in this study or its conclusions. Yet a confidence interval using the Fallujah data has been expressed (risk); moreover this CI is being incorrectly used to back into the confidence interval for excess death [the one that wasn't given out for the reasons we agree upon.] It has led to much confusion among the Tim Lamberts and the Kevin Donahues, deceptively leading them to a "minimum value" that is not justified by anything in the paper itself. The "minimum" expressed in the paper is 8,000. It is not 100,000. It is not 1.6 times anything. The CI for excess death uses a different method from the CI for risk; its lower bound is independent of the lower bound for risk. Tim continues to ignore this point even though it's laid out explicitly in the report itself.
I'll up the reward to $10 USD (or the currency equivalent of your choice) and an elaborate personalised apology for anyone managing to extract a CI for excess death from any of the authors themselves, whose 95% lower bound lies above zero. This should be easy money for someone.
Since Pascal's comments seem a bit incoherent to me, it may be worth setting out what I think he is trying to say.
From a purely statistical viewpoint, the purpose of the Lancet study is to estimate two population parameters: relative risk and excess deaths. These are unknown constants. Obviously there is a positive relationship between their possible values.
A sample is used to calculate two estimates, using two different estimation procedures. Let X denote the point estimate for excess deaths and R the point estimate for relative risk. The researchers also generate cumulative functions in order to get confidence intervals. Denote these functions X(.) and R(.), respectively.
Pascal's Conjecture is: Although X(1.6)=0.025, R(0)>0.025.
Two questions arise:
Is Pascal's Conjecture correct? It may be, but I see no reason whatever to believe it.
If Pascal's Conjecture is correct, does it matter? No. The sensible interpretation of such a result is that the excess-death estimator returns a freak result in the presence of an observation like Fallujah. As I said before and as others have said in different ways, true mortality doesn't fall when battlefields are included in the total, whatever the computer may say.
That should read:
Pascal's Conjecture is: Although R(1.6)=0.025, X(0)>0.025
The sensible interpretation of such a result is that the excess-death estimator returns a freak result in the presence of an observation like Fallujah.
This is circular or selective reasoning, more "appeals to intuition." The Fallujah cluster itself is a freak result. The sensible interpretatation is that the result itself cannot be used to calculate any meaningful confidence interval. Deciding which CI to report and which to omit should not be based on what might "discredit the survey" or what might appear "too wide" or what might nullify the stronger "excess>0" claim. Neither confidence interval belongs in the report, or both do. Kevin at least acknowledges that X and R CIs are not functionally related. I hope he recognizes that using 1.6 to make any statement re: excess deaths is fallacious, and therefore that claiming "at least 100k" on this basis is false.
Apropo of nothing:
Pascal:
The Fallujah cluster itself is a freak result.
We don't know that. There may be many areas in Iraq which have experienced similar violence.
Kevin at least acknowledges that X and R CIs are not functionally related.
If by that you mean that they are stochastically independent then I don't see it that way at all. The two CIs are functions of the sample data. To be more precise, there is a mapping from the set of possible samples to the X(.) function and another to the R(.) function. In repeated sampling they would vary in a systematic way.
I hope he recognizes that using 1.6 to make any statement re: excess deaths is fallacious....
Any statement? That is much too sweeping a claim. There is some true number for relative risk and excess deaths. The higher the former is, the higher the latter is. If true R>1.6 then true X>0. Also, if the R(.) function is trustworthy, either true R>1.6 or a 1-in-40 fluke has happened.
Let's reduce the scope for misunderstanding. That comment should have read: To be more precise, there is a mapping from the set of possible samples to the set of possible X(.) functions and another mapping to the set of possible R(.) functions.
Sorry to be so pedantic, but it seems to be warranted.
There may be many areas in Iraq which have experienced similar violence.
The fallujah cluster is a freak because its conclusions even regarding fallujah are wildly unlikely. This is why the "intuitive argument" grates. You are making intuitive appeals on the basis of an data that suggests 200k excess deaths occurred in Anbar alone. (If I weigh myself five times, and the results are 180lbs, 182lbs, 182lbs, 184lbs, and then 2,000lbs, do I conclude per the final measurement that intuitively, I probably weigh more than 182? Is the final measurement permitted to inform any calculation at all? This is the scale of the outlier.)
I am questioning the process that led to an expression of an IFC CI for risk and not for excess death. Since you are attempting to use one to "back into" a meaningful boundary for the other, this objection is reasonable. If true R>1.6 then true X>0 does not address the CI question, since we know the CIs to have been calculated using different bootstraps (one using linear regression, one loglinear).
I'm not certain how the risk CI was obtained, or why loglinear regression was applied in that bootstrap versus linear regression for the other. I am quite certain that using the outcome of the risk CI to make any deductions regarding the other CI is incoherent.
If the lower bound of the study-wide risk is 1.6, the authors should have no problem revealing the excess deaths associated with this figure, calculated in a manner consistent with that of the ex-Falluja excess death interval. Their failure to represent this number invites guesswork and phony deductions, and in this regard is irresponsible.
I now see I am in the presence of a clown, so I forgive you, Pascal, for missing the first sentence of the "findings" section before and will try the tutorial one more time. I won't be responding to personal remarks, though I might chuck out a few of my own.
Try 32 numbers between zero and three, averaging 1, with a standard dev of .5 or so. Then throw in a 66; the standard deviation jumps by a factor of 7. The 95% confidence interval now extends well below zero.
No it doesn't. You're confused about what is meant by "the confidence interval". Samples don't have confidence intervals; distributions do. Now, we can choose from a number of distributions when we are trying to construct a confidence interval; for example, we can use the empirical distribution suggested by the data, or we can make a parametric assumption. If you're talking about the empirical distribution, then the 95% confidence interval is the 5th percentile of the dataset, which (by construction in this example) is not below zero.
What you appear to be talking about is the 95% confidence interval based on the parametric assumption that the data were normally distributed. But that would be a stupid parametric assumption to make given the data observed; as Eli points out above, it implies that we are equally likely to see a similarly sized negative outlier and the data do not support this claim.
Therefore the correct thing to do is bootstrap the confidence intervals from the data. If you don't understand this point, you don't understand why the bootstrap was used and you don't, therefore, understand how the condfidence intervals were constructed.
This obvious and simple point has been made again and again, to quizzical stares from the crowd of supposed stats experts
Yes; we can't believe that you keep on making this point despite having had it explained several times that it's wrong.
You are making intuitive appeals on the basis of an data that suggests 200k excess deaths occurred in Anbar alone
No; not in Anbar. In the 3% of Iraq represented by the Fallujah cluster. This is not the same thing; please try to be clear.
Kevin: I think that "Pascal's Wager", as one might put it, can only be won if there was material correlation between the uncertainty about the prewar deaths estimate and the uncertainty about the relative risk ratio. As far as I can see, the only way one might get this would be through some version of the "Lying Iraqis" theory; that those lying Iraqis hated us enough to invent deaths of family members postwar and ignore them prewar. This might do it, although quite obviously a) there is nothing in the data to support a positive estimate of correlation between the uncertainties in this way, b) there is decent evidence (the death certificates) to suggest that the correlation is zero as it should be and c) anyone making this critique would do a lot better to just say that they believed in the "lying Iraqis" data problem than to play silly-buggers about confidence intervals.
More generally, though, I'd invite anyone looking at this debate to have a close look at what Pascal is trying to achieve with his current line of argument, because it's a textbook example of the lack of good faith that is so distressingly common on the denialist side.
Pascal is now expending massive rhetorical effort in trying to push 5% of the distribution for excess deaths below zero, ignoring what might be happening to the other 95% of the distribution. In the context of risk management comittees and regulatory meetings, where you also often get junior ticks trying funny things with parametric assumptions to squeeze the 95% "value at risk" under some limit, I tend to call this behaviour "percentile fetishism" and suggest that it is the statistical equivalent of coprophagia. In the context of the Lancet debate, I'll stick to calling it "Kaplan's Fallacy". (There is also a kinship resemblance to the Steven Milloy relative risk ratio criterion; all "sound science" hacks are cut from the same basic template).
s now expending massive rhetorical effort in trying to push 5% of the distribution for excess deaths below zero, ignoring what might be happening to the other 95% of the distribution.
I know exactly what it does to the rest of the study; it triples the central estimate. It claims that 200k excess deaths occurred in the 3% of Iraq represented by the Fallujah cluster. It discredits and undermines the study completely. Which is why I prefer (as should you) that it be ignored completely, not partially and selectively.
I am discussing a paradox contained in two distinct lines of the study's defense ("the CI doesnt contain zero"-good and "We're not even counting Fallujah/at least 100k"-bad /stupid). Your point (and Eli's) regarding the distribution from which the Fallujah cluster was drawn is the same as mine and is made to illuminate the hazard of expressing a Fallujah inclusive confidence interval ANYWHERE within this discussion; I really wish you would stop claiming it to be something I am ignoring or trying to deny. This well understood assumption should be causing you to question the Fallujah-included CI associated with mortality risk, and the several false interpretations of this variable, not getting in my face. If I were looking to defend the study, I really wouldn't want to discuss Fallujah in terms of any confidence intervals or the number it adds to the mean. You as a master econometrist should be taking Tim and Kevin to task for their indiscretion, not strutting around like an ass complaining about bad faith. Bad faith is claiming that Fallujah described a realistic result, and incorporating this result selectively in your depiction of the study conclusions. Bad faith is pretending not to know how to read, or why an example has been drawn, or why one line of argument is coherent, cogent and responsible, and one is not.
Please explain if you can why the IFC confidence interval applied to risk of death is expressed as a finding of the study. If you can, may we hear your opinion on whether this allows us to know any characteristic of the CI for excess deaths? If so, what are the bounds of this CI? If you don't know, I'll understand completely.
Thanks in advance.
Reminder: this portion of the discussion revolves around whether "at least 100k" is a fair portrayal of the study's conclusions, based on the Fallujah cluster. This is what you are defending.
Pascal ol fella, go read the latest on how many houses in Falluhjah were destroyed and then try and tell me how the Fallujah cluster is typical of all of Iraq. Basically Fallujah is not on the far end of some normal distribution of death and destruction in Iraq. It and a few other cities are off by themselves.
Given these FACTS anyone who tries to include the Fallujah cluster into some distribution of excess deaths in Iraq is basically trying to average down the damage done there.
An amusing form of denial to be sure.
It claims that 200k excess deaths occurred in the 3% of Iraq represented by the Fallujah cluster.
Better expressed: it claims that 200,000 excess deaths occurred among the 739,000 people represented by the Fallujah cluster.
Eli, for the millionth time, I'm not trying to average anything down. Your points about the distribution are the same as mine. Which is why the representing the IFC rate CI is so ill-advised. It clearly has invited all kinds of irresponsible guesswork and bad logic regarding the excess death CI. The study's authors could clear this up pretty easily, and there's a crisp 10 bucks in it for you if you get them to do it. No risk to you! How could this be a bad faith argument when I am offering you cash rewards to advance it?
As a non-stats expert I was initially surprised by the claim that including Fallujah might widen the uncertainty so much it'd make a decrease in the death rate possible. But then Heiko (of all people) and others made me realize this was only on the assumption that what happened in Fallujah was an extreme example of what was happening in the rest of Iraq. This is transparently not the case.
Therefore the non-Fallujah data gives us the 8000-194,000 increase in deaths, and the Fallujah cluster, by showing that a few places in Iraq suffered much higher levels of destruction, suggests that the number is more likely on the high side. This is easy to understand. People are still free to think that the actual death toll is on the low side, based on whatever else they think they know about Iraq, but not because of Pascal's statistical argument.
Donald is entirely right, and I note that Pascal, while continuing to chuck around silly personal insults, hasn't quite got round to mentioning what I thought was a quite lucid explanation of the importance of keeping in mind the difference between the empirical distribution of the data on the one hand, and some form of unrealistic parametric assumption about that data on the other. (here's the link; it wasn't too long ago). The Lancet guys, quite correctly, printed the CI for the relative risk in their complete dataset, said that Fallujah was an outlier, explained why and printed a CI for the dataset ex-Fallujah.
This gives us (as Eli entirely correctly says) an estimate of 98K excess deaths in the 97% of Iraq represented by the non-Fallujah clusters. We then have a potentially very significant number of excess deaths in the 3% of Iraq represented by the Fallujah cluster. The point estimate would be 200,000; this is certainly not a ridiculous number given that Fallujah was a town of 250,000, only about 10,000 can currently be accounted for, and it represents a number of other high-violence regions. But, it is only one cluster of 33 households, so it is difficult to make any statements with confidence about the relative risk ratio in Fallujah-like clusters. However, it is quite clear that the data that supports the claim "at least 100K", since the point estimate would be 10130K if Fallujah was actually no more violent than the rest of Iraq and all the evidence we have (sample and non-sample evidence) suggests it was much more violent.
Pascal seems to believe that the point estimate of 200,000 which might have been extrapolated from the Fallujah cluster (but was not) "discredits the study completely". There is no reason for asserting this other than that's what Pascal wants to be true (one good way of spotting a hack is that he starts pretending to very exact statisical argumentation when he thinks it supports him but then retreats to hand-waving "this is obviously wrong" when it doesn't). His analogy with a weighing machine is really, really silly; your weight doesn't have much variance from day to day, but the count of deaths in different regions of Iraq clearly has a lot of variance.
People are still free to think that the actual death toll is on the low side, based on whatever else they think they know about Iraq, but not because of Pascal's statistical argument.
Donald, I am not saying Fallujah makes a lower number "more likely" than a higher number, just that it likely makes a lower number more likely than if Fallujah were excluded altogether. I am saying it was excluded or the other CI included for a bad reason. I am agreeing with dsquared that all CIs drawn from all heavily skewed distributions are to be viewed with great suspicion, and therefore questioning the usefulness of a CI surrounding Fallujah-inclusive "risk of death." Also its use to justify the false claim that the study found "at least 100,000" excess dead or anything similar.
The point estimate would be 200,000; this is certainly not a ridiculous number given that Fallujah was a town of 250,000, only about 10,000 can currently be accounted for
Even the study authors have noted that 200,000 is not a reasonable claim. The claim about discrediting the survey comes from Roberts himself, who said that including this cluster would have discredited the survey and undermined its validity. I can't understand why you don't take his word for it. Maybe it has something to do with this data informing the reported risk of death and its CI? Dsquared, why do you think no confidence interval was reported for excess deaths? Do you have any thoughts on whether the CI for Fallujah-inclusive risk a.) is a useful or accurate measure or b.) can be used to back into a CI for excess death as Tim claimed earlier?
I'm cool with empirical distributions of data, in which case I wonder why you all see fit to throw around interior CIs with any confidence at all, seeing as nobody knows the shape of the bootstrapped probability distribution here.
one good way of spotting a hack is that he starts pretending to very exact statisical argumentation when he thinks it supports him but then retreats to hand-waving "this is obviously wrong" when it doesn't
EG: your embarrassing attempt to use Fallujah as a surrogate for all other heavily bombed regions in Iraq, sneaking it into the final estimate under your shirt with 57 varieties of "intuitive appeals" that to me are anything but. (I love it when MBA back office quants, not satisfied with their game of dress up as epidemiologists, play at military analysis as well.) To which my own "intuitive argument" was a direct response, as I made clear. Read slower and more carefully.
Eli, if you are arguing with me now, you are defending the "at least 100k" characterization of the study, on the basis of the "semi-excluded" outlier. I am simply advocating a consistent approach to the presentation of such data.
While the original topic of this thread may have been the use of the phrase "at least 100k", what got me interested was Pascal's unshakeable conviction that the unpublished Fallujah-inclusive excess-death CI extends below zero. Pascal hasn't given us a serious reason why he believes this, just some assurances about his experience with data and the usual insinuations about Roberts et al. The theory advanced above by dsquared, that it is just a fancy version of the "lying Iraqis" story, is probably the best we are going to get.
Enough already. The relative-risk CI has been fairly admitted in evidence and there is no reason to avoid using it whenever a reasonable deduction can be made from it. The statement that excess deaths can only be less than zero if a 1-in-40 fluke has transpired is such a deduction. Pascal may as well get used to the fact that his conjecture is just that and nothing more. Maybe Andrew Wiles will undertake the task of proving it. He did it for Fermat, why not Pascal?
The relative-risk CI has been fairly admitted in evidence and there is no reason to avoid using it whenever a reasonable deduction can be made from it.
Testimony admitted into evidence is subject to cross examination. No one here seems eager to take up that chore. I wonder why? Dsquared has yet to weigh on the aptness of discussing a confidence interval for "risk of death" distribution, fallujah inclusive, although he has opined compellingly and at length on the asymmetry and non-parametric character of the 'excess death' distribution. Yet he says nothing as ambitious (and vain) attempts to map functions between two different CIs take place under his nose, and as a CI is implied for data for which he denies such a property exists. This is bad. It detracts from the credibility of the report, and its advocates. A straight answer is needed here. If the report contains as part of its conclusion a boundary or range whose corrollary is a higher (or lower) bound for the excess deaths, we should certainly be allowed to know what this value is. "At least 100k" is wrong, but licensed and bonded statisticians & epidemiologists continue to use the expression. They use the IFC data and one of its risk intervals as the basis.
Pascal,
Never mind the stream-of-consciousness stuff. If you know of a problem with the Fallujah-inclusive relative risk estimate, tell us what it is.
As to "ambitious (and vain) attempts to map functions between two different CIs", these attempts exist only in your imagination. When I say that (absent a 1-in-40 fluke) the relative-risk CI implies that true excess deaths are positive, I am saying nothing whatsoever, good or bad, about the unpublished excess-death CI.
An obvious point for dsquared to consider: assuming he is earnest in wishing to retain Fallujah data as part of a separate distribution within the survey results even that distribution must contain a confidence interval. What would its bounds be - lower and upper? We know they would be very wide indeed, but just how wide? Conveniently unknowable.
If you know of a problem with the Fallujah-inclusive relative risk estimate, tell us what it is.
Kevin, my problem is that is has been calculated and expressed as a result of the survey and then misused to make parallel inferences. I (and dsquared and Eli Rabbitt and maybe others) think that outlier data should be considered apart from core study data. We don't think confidence intervals alluding to "bi-modal" or skewed or two utterly distinct data sets have informational value. We (or rather only I) feel the "1.6...R(0)>.025" inference invalid, as it employs the lower bound of one such CI. This series of missteps have been used to justify a misrepresentation of what the study demonstrates ("likely >100k excess deaths".)
What would its bounds be - lower and upper?
The answer to that question is contained in the survey, but I think that a short read of the survey would have a salutary effect on you, so you can look it up yourself.
I (and dsquared and Eli Rabbitt and maybe others) think that outlier data should be considered apart from core study data. We don't think confidence intervals alluding to "bi-modal" or skewed or two utterly distinct data sets have informational value.
No. There is all the difference in the world between thinking that a single "95% confidence" figure isn't the be-all and end-all, and thinking that it has "no informational value". The difference is, to be blunt, the difference between an irritating wet-behind-the-ears graduate trainee or internet arguer on the one hand, and someone who has some experience in getting his hands dirty with live data on the other.
Pascal, include me out
These "angels dancing on pin head" statistical wranglings have gone too far enough already! What is needed is some anecdotal realism, based on comparitive history. In GW I, Frontline reports, the US military, in the course of operations, inflicted around 25,000 fatalities on Iraqi persons and personnel:
................
There is some evidence that casualties in the Iraqi army were lower. In 1993 John G. Heidenrich, a US intelligence analyst, published "The Gulf War: How Many Iraqis Died?" in Foreign Affairs, in which he estimated that
................
Whatever the correct figure (my money is on Heidenrich) it is the case that most of these casualites occurred during formal encounters on open ground with military installations or institutions. In those days the US military's aresenal of weapons were less deadly and less accurate, meaning a lower kill-to-ordinance ratio.
In GW II, by contrast, has gone on for two years, in a mostly urban environment, using more accurate and deadly weapons against mostly irregular soldiers, in the context of a severely weakened community service infrastructure. So one would not be surprised, given the context and duration of military activities, to see that overall Iraqi casualties around one order of magnitude greater than the top-range of Heidenrich's estimate (~10,000). This is consistent with Lancets estimation of ~100,000 excess Iraqi deaths .
As reluctant as I am to end my self-imposed exile from the increasingly tenuous, not to mention tendentious, arguments on this blog (I'm entirely in agreement with Jack Strocchi on the matter of pinheads), I have to say that an expert who estimates that under 100 civilians were killed in "GW I" doesn't seem any more firmly grounded in reality than some of the participants here. More were killed in a single attack on a Baghdad air-raid shelter.And BTW the Lancet's estimate applies to 18 months, not 24.
Tendentious argumentation indeed. Argumentation from Googlers, to boot.
Google doesn't have a "context" or "wisdom" button, making d^2's assertion above especially cogent. My clue that someone doesn't know their subject is the tiny dissection of little points, then blowing them up to speak to the larger topic. Rarely works, certainly not here.
Best,
D
My clue that someone doesn't know their subject ...
More tedious authority-mongering. Dsquared's wisdom on matters econometric is known to me already, thanks, as is his charming debating style.
who has some experience in getting his hands dirty with live data on the other.
Is this in a "risk management" or "regulatory" context I wonder? This conversation has made me think a bit more of the former ;) Usually the mindset most likely to poo-poo the relevance of outliers can be found in government work.
'bit more of the latter', dammit.
I'd like to address one of Kevin's points, namely that many people supporting the invasion thought it would result in a negative excess death figure, while many supporters now say "well, war kills, so what else would you expect".
These statements aren't mutually exclusive. It is, in fact, quite reasonable to expect a short term increase in death rates. Virtually no matter how well military action and regime change are executed some disruption of services and killing is to be expected.
This would be counterbalanced by longer term gains. Which throws up the question of how long a time period to consider for excess death? A week, a month, a year, a decade, several decades?
And what to compare with? It's pure guesswork as to what would have happened, if Saddam had stayed in power. Another brutally suppressed uprising of the Shia? A successful orange revolution?
More tedious authority-mongering. No one is appealing to authority. Someone is pointing out that most people using your nit-pickicious debating style can't identify a hole in the ground, hence the tactic being a clue.And hence the spammed comments on this topic on poor Tim's site.HTH,D
One issue I have with discussions of "excess death" is how it effectively equates car accidents with suicide bombings or the killing of an insurgent.
If it really made so little difference, why concentrate on Iraq, where so little of the world's excess death is occurring? A million traffic fatalities, around 10 million children dying needlessly every year, when they could easily be saved with today's medical technology ...
For me, it does make an enormous difference, whether 100,000 Iraqi civilians have been killed by American bombs, or wether 33,000 insurgents got killed by American bombs, and 33,000 civilians got killed by insurgents, and 33,000 in additional traffic accidents.
Which gets me onto what the study has to say about bombing deaths.
If we assume that 2/3 of the bombing deaths are in high bombing intensity clusters (say 1-99 deaths per cluster), and 1/3 in low intensity clusters (say 1-33 deaths per cluster), and we assume 100,000 bombing deaths, a few simple spreadsheet operations show that in about one third of samplings a Fallujah type cluster would turn up. If it was included, it would virtually always give an overestimate of the true number (assumed to be 100,000). If it was excluded, it would virtually always give an underestimate of the true number. Typically it would give a number in the low ten thousands.
Basically, this is an indicator that if the Fallujah cluster really fairly represented "high intensity bombing" areas, 100,000 bombing deaths would indeed not seem all that far fetched.
To reject that requires assuming that the Fallujah cluster is likely representative of a smaller (or as I think much smaller) proportion of Iraq.
A few simple spreadsheet operations also show that getting three bombing incidents outside of Fallujah and 6 deaths with just 3000 total bombing deaths across the country (all in low bombing intensity clusters, ie 1-3 deaths per cluster) is very unlikely to occur due to mere random sampling error.
Which either indicates that the true number is indeed higher, or that such a freak sampling error occurred, or that the data are false in some fashion (not correctly sampled, lying, misunderstandings etc...).
It should say 1-3, not 1-33