Smoking Trials Again

Recently I talked about checking on smoking skeptics.  I described three studies:

  1. A randomized trial of 1400 high risk smokers.  After 10 years one half had half the smoking rate of the other, and after 20 years it had an insignificant 7% lower mortality (13% less heart disease, 11% less lung cancer).
  2. MRFIT randomized multifactor trial of 8000 smokers.  After 6 years one half quit 49% (vs. 29%), and after 16 years had an insignificant 6% lower mortality (11% less heart disease, and -15% less lung cancer).
  3. A randomized multifactor trial of 1200 high risk men.  After five years one half reduced smoking by 3/8 (vs 2/9), but had twice the mortality (10 vs. 5 count).

I’ve now had time to look over seven more studies:

  1. A randomized trial of 6000 smokers with “asymptomatic airway obstruction”, i.e., weak lungs. (HT Karl.)  After 5 years in two-thirds, 22% (vs 5%) stopped smoking, and after 14.5 years they died a (3% level significant) 15% less (20% less of heart disease, 15% less of lung cancer, and 50% less of “respiratory disease other than cancer.”) (More details here, which I don’t have.)
  2. WHO collaborative multifactor randomized trial of 61,000 men.  After six years one half had 2% fewer smokers, 7% among highest risk men, giving an insignificant 5% lower mortality (7% in heart disease).
  3. Gotenborg multifactor randomized trial of 30,000 men.  After ten years one third had 9% fewer smokers (32.5 vs. 35.4%) than the other two thirds, and an insignificant 2% lower mortality (0% heart disease, 15% cancer).
  4. Norwegian multifactor randomized trial of 1200 men.  After five years one side had 1/8 less smoking, and after 28 years it had 46% more mortality (95 vs 65 count).
  5. Oslo mulitfactor randomized trial of 1200 men.  After 8.5 years one side had 45%(?) less smoking, and 40% less mortality (19 vs. 31 count).  (This just from abstract; anyone have the paper?)
  6. A non-randomized study of 1600 men over 26 years. Initial lung quality was unrelated to mortality for non-smokers, but high smokers with initially bad lungs died 62% more than initially good lungs.
  7. A non-randomized AER ’06 study of WWII vetrans.  Its key “identifying assumption is that cohort and age effects in the smoking equation are the same for men and women” and that the entire increased mortality of WWII veterans is due to their smoking more. (HT Alex T.)  It finds “a nonveteran average annual mortality rate of 13.1 per 1,000 men and a veteran … rate of 16.6″ (1.2 vs. 2.2 for lung cancer), suggesting “36 to 79 percent of the excess veteran deaths due to heart disease and lung cancer are attributable to military-induced smoking”.  Since heart disease and lunch cancer were 38% of deaths, this suggests ~4-12% higher smoking mortality.

OK, so how best to summarize this evidence?  Based on study #4, I tentatively estimate smoking raises mortality for folks with bad lungs, about 10 to 25% of folks, by 50-100%.  (This affect appears to not work mainly via lung cancer.)  This is supported by study #9 and could explain a 5-25% overall smoking mortality increase.

In the rest of the studies, if we assume the entire effect seen was from smoking, we can collect smoking mortality affect estimates.  Setting aside #8, as I haven’t read the paper, #1 had the biggest change in smoking rates, and suggests a ~20% mortality.  The next biggest change was #2, and suggests ~30% mortality.  Study #6 had the next less change, and suggests ~22% mortality.  The rest were all across the map, as expected from their small count and change.

So, we seem to see a 50-100% smoking mortality increase on bad lungs, which predicts a 5-25% overall smoking mortality increase.  If we attribute to smoking the full benefit seen in our three most relevant multifactor randomized trials, we get crude smoking harm estimates of 20,22,30%.  And if, from study #10, we attribute the entire higher mortality of WWII veterans to their smoking more we get ~4-12% mortality effect.

Bottom line:  a randomized trial suggests a large smoking harm on bad lungs, which can explain the entire apparently average smoking harm seen elsewhere.  My best guess: smokers die ~10-30% more on average, living about 2-6 months less, but there’s much less net harm to strong lung folks.

Added 10a: Wikipedia says

Male and female smokers lose an average of 13.2 and 14.5 years of life, respectively. .. The risk of dying from lung cancer before age 85 is 22.1% for a male smoker and 11.9% for a female current smoker, in the absence of competing causes of death. The corresponding estimates for lifelong nonsmokers are a 1.1% probability [20 times less] of dying from lung cancer before age 85 for a man of European descent, and a 0.8% probability [15 times less] for a woman.

Other sources mention risk factors of 15, 23 or 100. Such figures are common and, it seems, rather misleading. The above studies clearly suggest that the causal effect of smoking on mortality, even for lung cancer, is much less than the factors of 15+ often thrown around.

GD Star Rating
a WordPress rating system
Tagged as:
Trackback URL:
  • http://rationalitate.blogspot.com Stephen Smith

    You seem to be into drug myths (date rape drugs, nicotine, etc.), so you might enjoy this one: heroin, in isolation, is virtually impossible to overdose on. Most deaths attributed to heroin use are actually due to concurrent use of heroin and other CNS depressants, like as alcohol and benzodiazepines.

    On its surface, this would seem to be just an interesting harm-reduction strategy – if people choose to take heroin in combination with other CNS depressants (which they clearly often do, owing to the number of heroin overdoses), the simple fact that heroin alone doesn’t kill doesn’t undermine the case for prohibition. However, there has been some behavioral econ research done on heroin addicts that shows that alcohol and benzos are second-best substitutes to heroin, and their value to heroin addicts drops when the price of heroin drops. This implies that “overdose” deaths are collateral damage from policies like prohibition which raise the price of heroin relative to legal (sort of) drugs like alcohol and benzos.

  • diogenese

    Robin — smoking CAUSES “bad lungs”. Chronic Obstructive Pulmonary disease (COPD) — like lung cancer — is primarily a disease of smokers. (although not as heavily skewed, I would guess ~ 4 smokers, for 1 non-smoker). Mechanisms by which smoking induce COPD are well studied. There is obviously a gene-environment interaction (not everyone who smokes, develops severe COPD — but people who otherwise would have had ok lungs, develop COPD when exposed to toxins, such as cigarettes).

    The more half-baked posts you have on this topic (such as very poorly cherry picking evidence in favor of your childish need to be “contrarian”) — the more you are going to permanently discredit yourself on any topic in health care.

    This post again is EXTREMELY misleading, as you are selectively ignoring the very strong observational data.

    • http://hanson.gmu.edu Robin Hanson

      I’ve described ten studies so far. I asked folks before to point me to the best study they thought I should consider. Which exact study do you suggesting I’m purposely ignoring?

      • diogenese

        Robin — you have not cited ONE review article in any of your threads. Have you even bothered to look at one? There are thousands of studies on smoking —- randomly selecting 10 studies without rhyme or reason doesn’t make an argument. For a topic this BROAD — you have to start at a review article to find important studies (then of course you can argue against them with other studies you find yourself).

        Studies:
        1. Doll R, Hill AB. Lung cancer and other causes of death in relation to smoking; a second report on the mortality of British doctors. Brit M J (1956) 2:1071–1081.

        2.) Schwartz D, Denoix P. L’enquette francaise sur l’etiologie du cancer broncho-pulmonaire: role du tabac. La Semaine des Hopitaux de Paris (1957) 33:424–437

        3.) egi M, Fukushima I, Fugisaku S, et al. An epidemiological study on cancer in Japan. Gann (1957) 48:63.

        4.) Dorn H. Tobacco consumption and mortality from cancer and other diseases. Acta Unio internat contra cancrum. In press.

        5.) Hammond EC, Horn D. Smoking and death rates – report on forty-four months of follow-up of 187,783 men. JAMA (1958) 166:1159–1172. and 1294–1308.

        Historical prospective and review of data can be found @
        http://ije.oxfordjournals.org/cgi/content/full/38/5/1175#B9

        Or start @ Surgeon Generals 2004 update report on smoking risk.

        BTW — Obviously as this thread indicated you somehow had no idea what COPD was, it seems equally clear you don’t know what a dose-response curve is. I don’t feel like its my job to educate you on pack-years and risk of disease.

      • http://hanson.gmu.edu Robin Hanson

        diogenese, this is why experts use big words – I say “weak lungs”, a phrase my readers will understand, and you say I know nothing because I do not call it “COPD”. (Apparently you also want me to throw out phrases like “dose-response curve” too.) I’ve focused initially on recent and randomized data, so naturally I haven’t been discussing your four non-randomized studies from the 1950s. Your “review” article is by Cornfield, one of the main partisans in this dispute, so it is hardly neutral, and it doesn’t mention randomized trials of any sort.

      • diogenese

        Robin — dose-response curve is a term in any introductory text on epidemiology, statistics, or causal inference. It basically means the more you smoke, the higher your risk — people who smoke less have less risk. I would say its an “expert term” as much as “opportunity cost”, “supply-demand”, or “variance” are “expert terms”

        FYI I’ve cited the original studies — you can find their modern updates yourself with google. Its already clear you can’t be bothered to look at a review article, you can look some thing up yourself.

        HEH this point is utterly classic — you should be EMBARRASSED. Your “review” article is by Cornfield, one of the main partisans in this dispute, so it is hardly neutral, and it doesn’t mention randomized trials of any sort.

        That’s after you post studies funded by the Tobacco Industry and say they are “professional”. Seriously you should be ashamed of yourself.

        I guess reviews paid for by the Tobacco industry can be taken seriously, but those that believe smoking is linked to cancer and increased mortality aren’t to be taken seriously.

        I think the first step to “overcoming bias” — would be to get over your childish need to be contrarian.

      • retired phlebotomist

        COPD is not obscure jargon for “weak lungs.” It’s chronic bronchitis and emphysema– the 4th leading cause of death in the US.

        I do appreciate your dogged pursuit of your position. I recommend you add this topic to your health econ class. It has certainly helped me evaluate how committed you are to your putative quest for truth.

        Glad you didn’t quit blogging back a year back when you kept bringing it up. Would’ve missed out on near v far and dreamtime.

        But you’ve definitely dipped permanently below the “high status” meridian on this one, so my desire to affiliate w you is at an end. ;)

  • Eric Johnson

    The non-RCT studies cited seem to use strong assumptions that would warrant dumping them.

    I somewhat agree with Diogenese. The (observational) 23x risk ratio for lung cancer still seems to clash with your conclusions. This number may be observational, but it is certainly quite simple epistemologically; it uses (or at least raises the question of) only one weak assumption. And that is that “smoker types”, if they never actually smoke, dont have *too* much higher a risk of lung cancer. And even if their risk is putatively higher by 2 or 3-fold, why then one can correct quite exactly for any falsity in this assumption: 23/3 is still a bit larger than unity!

    Whether the 23x has been confirmed several times I cant say. I got it from the usual anti-smoking literature, and I know at least one study cited in Wik is congruent. I have seen nothing contrary.

  • Eric Johnson

    If the observational risk factor were a more accustomed 2.5x or 1.7x, I would of course say “dump it” if you have RCTs to look at. Dump it like a bag of dirt, and dont worry for a second. But what I’m saying is that youre in a different world when there is a 23x observational hazard ratio. I cannot snub 23x like I would snub a typical weak tea hazard ratio, where it is nearly ipmossible to rule out selection bias etc accounting for much of the ratio.

  • Bill

    Robin,

    How many deaths do you want to be responsible for if you create urban legends with bad, insufficient, and casual review of statistical studies?

    Put a number on it.

    And, post it on a mirror and look at yourself.

  • http://www.nancybuttons.com Nancy Lebovitz

    What are the odds of developing weak lungs from other causes after you start smoking?

    • retired phlebotomist

      What he is calling “weak lungs,” not being able to push a certain volume of air thru a spirometer within 1 second, is more commonly known as COPD (as diogenes noted above).

      COPD is the #4 cause of death in the US, and 80-90% who suffer from it are smokers.

      My takeaway from this is is two-fold:

      1) quitting smoking after a long-term heavy habit yields surprisingly little benefit.

      2) former cult members who advocate freezing noggins but drive ragtops are more likely to circle the wagons than admit they might be wrong, even as they write books about the irrationality of disagreement.

      • http://hanson.gmu.edu Robin Hanson

        The claim that “quitting smoking after a long-term heavy habit yields surprisingly little benefit” seems at odds with the fact that the strongest result we’ve seen in a randomized trial was of weak lung middle aged smokers who quit.

      • http://www.hopeanon.typepad.com Hopefully Anonymous

        “former cult members who advocate freezing noggins but drive ragtops”

        why discredit yourself like this?

  • http://hanson.gmu.edu Robin Hanson

    Eric, why can’t a combination of publication selection biases and smoker selection effects explain reported factors of 23? Why must that be mostly causal?

    Bill, I take responsibility for both the harms and the benefits of publicly looking at the data as honestly as I can. If you think I’m neglecting some important study, point it out to me.

  • Bill

    OK Robin, Here is your database of industry studies collected through litigation.

    http://www.tobacco.org/News/98minnesota.html

    • http://hanson.gmu.edu Robin Hanson

      How exactly is pointing me to a 70,000 word web page listing hundreds of media news stories on a lawsuit anything like pointing me to an important study?

      • Bill

        As stated below, this is the tobacco litigation database which contains industry studies on tobacco and lung cancer. It is true that it has the history of the litigation; it is also true it has the studies. Not mutually exclusive.

  • Bill

    Robin, What I linked you to was the database of litigation documents which include Tobacco industry studies collected from litigation. True, it also collected the news stories, but IT IS THE DATABASE OF LITIGATION DOCUMENTS AND SMOKING STUDIES.

    Also, what do you tell your son about smoking?

  • michael vassar

    There is obviously something VERY wrong with some method being used here, whether Robin’s or Eric et al’s. A more interesting discussion than condemnation would be to explain WHY the methodology he is using would be expected to not show the effect he is looking for, or alternatively, to explain, with actual probabilities, how the two methodologies could give such different results. If that can’t be done I think we should just pronounce science dead and bring out the healing crystals.

    • http://hanson.gmu.edu Robin Hanson

      I’ll bet that the relative risk ratio of owning an anchor for owning a boat is huge; people who own anchors are far more likely to own boats than non-anchor-owners. Yet I’ll also bet that randomized trials to give some folks anchors would not result in them owning many more boats. Now how mysterious is that?

      • michael vassar

        That example has the same logical structure as the one discusses, but it’s wildly implausible for smoking at the time studies first came out. For one thing, the fraction of smokers in the population is MUCH greater than that of anchor owners and was much more-so then than now. More importantly, no-one takes up smoking because they have lung cancer or know they are going to get it. Standard numbers amount to the claim that almost everyone who gets lung cancer is a smoker, and that this was even more true in the past. The causal assumption also strongly suggests that lung cancer rates should have fallen precipitously over the last few decades, as smoking became less common, independently from other cancer death rate changes.

        It seems more than slightly surprising that smoking could constitute an extremely strong predictor of some other behavior or cluster of behaviors which would selectively increase lung cancer risks by a huge factor while having only modest impacts on other risks while it is extremely a-priori plausible that smoking could damage the lungs and introduce carcinogens to them.

  • Eric Johnson

    Robin, I should check the literature and see if the 23x holds up in the hands of different workers. If all studies are around 17-30x, then a big effect from publication bias would seem unlikely. (Or — your detractors here might be thinking — *you* should check the literature! I myself am not quite that sharp-tongued, though I am pretty close! :) )

    Unfortunately I’m over-busy, but maybe I’ll dig into Pubmed later.

    But meanwhile, have you heard of the Bayesean paper by Ioannidis, “Why most published findings are false”, in PLoS Med? I suspect it is true, though I might think “30%+” are false, instead of “most”. I havent understood Ioannadis’ paper, but I have read probably 25,000 biomed abstracts and 5,000 papers, and its pretty amazing how many are contradictory.

    Anyway, I despaired of the math just glancing at it. But as a Bayesean and Physics PhD you might ingest it easily. It was much passed-around in biomed when it came out, and caused a big to-do.

    http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1182327/

  • Eric Johnson

    The other reason I dont think it’s publication bias is because Ive just never seen a risk ratio like that in medicine; not even close. Ive seen hundreds of studies (or abstracts at least) that I doubted on purely stats grounds. They tend to have risk ratios under 2.0 — sometimes a lot worse than that. And p values around 0.25 – 0.5. Of course the use of 0.5 as the maximum acceptable p val in medicine is ridiculous considering how much funding there is, how much publication bias there probably is, and how many crap results. I tend to call something I read a “maybe” unless it has p < 0.015 or so.

    If publication bias causes the 23x risk ratio, and there are rather a lot of reports, then we ought to *probably* see a spectrum of results down to 15x or 10x or so, even in studies with an ample sample size. If such papers are out there I'll reconsider.

    • http://hanson.gmu.edu Robin Hanson

      The risk ratios do vary quite a bit from study to study actually, down to 5x and lower. Take a look.

  • Curt Adams

    Actually several of these studies suggest quite large benefits. E.g. the Oslo study (45% less smoking and 40% less mortality) would crudely indicate essentially all deaths were due to smoking – by extrapolation if everybody quit the mortality would drop by 90% or so. Overall, the variability is quite high and the consensus benefit would be lower, of course.

    As a lot of people pointed out in the first round these studies don’t address the benefits of not smoking; they address the benefits of quitting after smoking for long periods. These studies would only show the 23x fold benefit of never smoking if all damage from smoking were fully reparable in the early part of the study (often a few years; sometimes just a few months). Genetic damage predisposing to cancer and lung structural damage are both essentially irreparable; atherogenic changes are reparable, partly, but only with very extreme diets which weren’t in these studies. So quitters will still have a greatly elevated risk, *due to smoking*.

    These studies are comparing groups of people who smoked a long time with other groups in which a small minority smoked somewhat less; even if the risk from smoking were a millionfold you’d only see a small difference.

    • http://hanson.gmu.edu Robin Hanson

      The WWII veteran study should include the effects of not starting to smoke, as should the randomized trials that include non-smokers, such as MFIRT.

      • Curt Adams

        But counteracted by some serious methodological problems – using time in the service as a proxy for smoking. That results in lots of smokers in the “nonsmoker” group. And, interestingly they estimate that smoking is causing 37-79% of excess deaths, which includes the range for conventional wisdom.

        MRFIT can’t say diddly about never-smokers. There are never-smokers, but they are selected for other serious health issues.

  • http://www.rationalmechanisms.com richard silliker

    Smoke-em if you got-em. Happy holidays everyone. Oh, keep your powder dry. Great fun,no?

  • Bob Unwin

    Glymour et al.’s “Causation, Prediction and Search” (p239) contains a fascinating discussion of the history of debates concerning scientific evidence for smoking’s effects on mortality. IIRC they discuss the skepticism of Fisher and others about smoking causing lung cancer, which were grounded partly inthe empiricist dogmas of the time but also on Fisher’s strong genetic determinism. They are critical of the medical establishment who pushed the causal hypothesis; they argue that the studies the establishment cited did not justify their conclusions (at least relative to their standards for statistical rigor). They say that what drove medical consensus (at least early on) was not unquestionable evidence of causation from the studies but rather the observational evidence + very strong priors for smoking being the cause rather than there being a common cause.

    IIRC they also mention intervention studies in which monkeys were made to smoke and had correspondingly higher lung cancer rates. This sort of study might be useful relative to the RCT’s discussed above in which it may be hard to measure how much those who reducing their smoking actually smoked.

    • http://hanson.gmu.edu Robin Hanson

      That is indeed a nice chapter. Your mention of animal studies inspired me to search for such. There are two reviews saying we do not find that smoking causes cancer in animals.

      • http://yudkowsky.net/ Eliezer Yudkowsky

        Well that got even me to sit up and say “What?” And the first thing I noticed about those studies is who does them:

        Lorillard Tobacco Co., Greensboro, North Carolina

        The tobacco companies have a reputation for playing very dirty with science. That was actually one of the first thoughts that occurred to me on reading your research summary – “Are results like this rather more likely to pop up with tobacco-company-funded research?”

      • http://hanson.gmu.edu Robin Hanson

        I will not dismiss these reviews because their authors work for a tobacco company. I’ve read them and they seem professional. There has been plenty of foul play on all sides in this history; if we used that to ignore folks affiliated with some side, we’d be ignoring all sides.

  • http://retiredurologist.com retired urologist

    I will not dismiss these reviews because their authors work for a tobacco company.

    IIRC, you had quite a different attitude about the pharmaceutical company-sponsored research used to justify FDA approval of SSRI’s for treatment of depression, as well as other pharmaceutical claims. Frankly, I’m stunned that you, with no expertise whatever in the fields of oncology or pulmonary physiology, would accept industry-sponsored information because “they seem professional” (to you). One of the first steps a rationalist would take to “overcome bias” is the elimination of conflict of interest.

    • drewster

      I think it maybe time to rename this blog “Confirming Bias”.

  • Pingback: Overcoming Bias : Animal Smoking Studies

  • Lexi

    A question re # 10, It is interesting to me that the military veterans have a higher rate of death, and I’m wondering if that is more likely to be attributed to the combination of smoking + trauma, as there are several studies that suggest that trauma reduces immune function and increases things like heart and cardiovascular disease? (skip to ‘findings’ http://xnet.kp.org/permanentejournal/winter02/goldtolead.html, while this article talks about child hood experiences that are adverse, it is likely that similar things are at play) (this article talks about PTSD and reduction in immune function http://books.google.com/books?id=DkXxsTuZbMEC&pg=PA531&lpg=PA531&dq=trauma+and+psycho-neuro-immunology&source=bl&ots=jupPtiucor&sig=2NWM0Lmx-935MQ8c1DPkUP1oAxo&hl=en&ei=a8g6S_3gB4LosQPiopTBBA&sa=X&oi=book_result&ct=result&resnum=7&ved=0CDEQ6AEwBg#v=onepage&q=&f=false)

    Anyway, I’m curious about what your thoughts are on those findings, and what role they may also play in the death rate of smokers?