WHY IS IT SO OFTEN REPEATED THAT SMOKING CAUSES CANCER? I’m not a tobacco user, so I’m not trying to justify my behavior. Has anyone here looked into the other things tobacco’s accused of causing or being “strongly” correlated with?
Smoking is not “accused” of being strongly correlated with negative outcomes. It is strongly correlated with negative outcomes, as a simple empirical fact. This is a statement about the joint distribution of the observed variables “smoking” and “negative outcomes”, and it has nothing to do with causal inference. I cannot even imagine a scenario where the statement “Smoking is strongly correlated with lung cancer” is false, short of a vast conspiracy among scientists and doctors
A slightly more interesting question is whether the correlation between smoking and cancer is due to causation. It is theoretically possible that an unmeasured confounder is responsible for the observed correlation. In fact, R.A. Fisher believed such a confounder was probably at work . One of the first uses of sensitivity analysis was to show how unrealistic Fisher’s claim was. A sensitivity analysis is essentially a thought experiment that lets you play around with how “strong” a confounder has to be, in order to account for the observed correlation if the causal null hypothesis were true. See http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2755131/.
In this case, I think any reasonable investigator who looks at the data and does some basic reasoning about possible confounders, will come away with a very strong posterior in favor of smoking causing lung cancer. However, the relationship between smoking and certain other negative outcomes is in some cases much more questionable, and it would not surprise me if publication bias accounts for many of the negative outcomes smoking has been connected to
Anders_H: “Smoking is not “accused” of being strongly correlated with negative outcomes. It is strongly correlated with negative outcomes....”
This is the opposite conclusion of the first citation I provided. And the second “in house” LW link asserts that in terms of decision making about smoking in light of whether or not it’s linked to cancer is about a 50⁄50 proposition.
Anders_H: ”...as a simple empirical fact.” This is a huge abstraction. Please clarify.
Anders_H: “This is a statement about the joint distribution of the observed variables “smoking” and “negative outcomes”, and it has nothing to do with causal inference.”
I understand that, but I’m not asking about that. I’m asking why the correlations are thought of as causes by reports on the relationship. And it is indeed an ACCUSATION commonly presented by the press, etc..., that smoking causes or is positively correlated to cancer. Furthermore, ccording to Hume, causal inferences are THEMSELVES observed by constant conjunction, implying we have know sure way of knowing what the relationship between causes and correlations is.
Anders_H: “I cannot even imagine a scenario where the statement “Smoking is strongly correlated with lung cancer” is false....”
Again, I refer you to the first citation, which also underscores the fact the line between “weak” and “strong” is done by fiat, another challenge to the so-called link between smoking and cancer.
The Japanese smoke more (if not the most) than most cultures yet are also one of the most healthy cultures. This goes to your “slightly more interesting question,” but it also goes the challenges of positively correlating smoking with “negative” outcomes. A further problem is that “negative outcomes” are normatively tied to cultural standards. Another problem is with average life expectancy comparisons, as they are to sensitive to outlier inflation.
This is the opposite conclusion of the first citation I provided.
Sorry, can you be more specific? Where does anybody claim that smoking is not strongly correlated with life expectancy?
And the second “in house” LW link asserts that in terms of decision making about smoking in light of whether or not it’s linked to cancer is about a 50⁄50 proposition.
The second “in house” link is a very simple thought experiment to explain the concept of confounding. It is meant as an example where evidential decision theory fails. In this situation, causal decision theory gives the right answer, it is certainly not a 50-50 proposition. Moreover, the correct answer within the thought experiment is that smoking does not cause cancer. This is because they postulated the existence of a deterministic confounder. This has no implications for whether or not such a confounder exists in the real world.
I ’m asking why the correlations are thought of as causes by reports on the relationship.
Because the confounders, ie the “smoking lesions”, would have to be unrealistically strong to fully explain the observed correlation between smoking and lung cancer. This is the part where I showed you the sensitivity analysis.
Furthermore, ccording to Hume, causal inferences are THEMSELVES observed by constant conjunction, implying we have know sure way of knowing what the relationship between causes and correlations is.
Of course we don’t have a “sure” way of knowing about causal relationships. But if you adopt “certainty” as your epistemic standard, you wouldn’t even be able to tell whether parachutes save lives in people who are falling from airplanes.
The Japanese smoke more (if not the most) than most cultures yet are also one of the most healthy cultures.
This is called an “ecologic” argument, and it is considered very weak. Note that your sample size is essentially 2, as the units you are making inferences about are countries, not individuals.
A further problem is that “negative outcomes” are normatively tied to cultural standards.
Now you’re just trolling… We’re talking about life expectancy, lung cancer, heart attacks etc here.
Continuing Causality Woes: Smoking and Lung Cancer:
Looking at:
http://lesswrong.com/lw/cc8/seq_rerun_changing_the_definition_of_science/
and
http://wiki.lesswrong.com/wiki/Smoking_lesion
Cross Referenced with Causation in the Presence of Weak Associations: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3024843/
WHY IS IT SO OFTEN REPEATED THAT SMOKING CAUSES CANCER? I’m not a tobacco user, so I’m not trying to justify my behavior. Has anyone here looked into the other things tobacco’s accused of causing or being “strongly” correlated with?
Background reading:
-Anything by David Hume
-Carl G. Hempel. Laws and Their Role in Scientific Explanation: http://www.scribd.com/doc/19536968/Carl-G-Hempel-Laws-and-Their-Role-in-Scientific-Explanation
-Studies in the Logic of Explanation: http://www.sfu.ca/~jillmc/Hempel%20and%20Oppenheim.pdf
-Causation as Folk Science: http://www.pitt.edu/~jdnorton/papers/003004.pdf
-Causation: The elusive grail of epidemiology: http://link.springer.com/article/10.1023%2FA%3A1009970730507
-Causality and the Interpretation of Epidemiologic Evidence: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1513293/
-Studies in the Philosophy of Biology: Reduction and Related Problems: http://books.google.com/books?id=NMAf65cDmAQC&pg=PA3#v=onepage&q&f=false
Do you believe the nutritional etiology of scurvy?
Yes.
Why? According to your blog, you don’t believe in RCTs, right? What do you believe in?
Start here. Follow the references (and the references’ references). If you are still not convinced then try here.
Did you follow the references I provided? Two of them are LW “in house” and the rest are superior to the ones you cited.
Smoking is not “accused” of being strongly correlated with negative outcomes. It is strongly correlated with negative outcomes, as a simple empirical fact. This is a statement about the joint distribution of the observed variables “smoking” and “negative outcomes”, and it has nothing to do with causal inference. I cannot even imagine a scenario where the statement “Smoking is strongly correlated with lung cancer” is false, short of a vast conspiracy among scientists and doctors
A slightly more interesting question is whether the correlation between smoking and cancer is due to causation. It is theoretically possible that an unmeasured confounder is responsible for the observed correlation. In fact, R.A. Fisher believed such a confounder was probably at work . One of the first uses of sensitivity analysis was to show how unrealistic Fisher’s claim was. A sensitivity analysis is essentially a thought experiment that lets you play around with how “strong” a confounder has to be, in order to account for the observed correlation if the causal null hypothesis were true. See http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2755131/.
In this case, I think any reasonable investigator who looks at the data and does some basic reasoning about possible confounders, will come away with a very strong posterior in favor of smoking causing lung cancer. However, the relationship between smoking and certain other negative outcomes is in some cases much more questionable, and it would not surprise me if publication bias accounts for many of the negative outcomes smoking has been connected to
Anders_H: “Smoking is not “accused” of being strongly correlated with negative outcomes. It is strongly correlated with negative outcomes....”
This is the opposite conclusion of the first citation I provided. And the second “in house” LW link asserts that in terms of decision making about smoking in light of whether or not it’s linked to cancer is about a 50⁄50 proposition.
Anders_H: ”...as a simple empirical fact.” This is a huge abstraction. Please clarify.
Anders_H: “This is a statement about the joint distribution of the observed variables “smoking” and “negative outcomes”, and it has nothing to do with causal inference.”
I understand that, but I’m not asking about that. I’m asking why the correlations are thought of as causes by reports on the relationship. And it is indeed an ACCUSATION commonly presented by the press, etc..., that smoking causes or is positively correlated to cancer. Furthermore, ccording to Hume, causal inferences are THEMSELVES observed by constant conjunction, implying we have know sure way of knowing what the relationship between causes and correlations is.
Anders_H: “I cannot even imagine a scenario where the statement “Smoking is strongly correlated with lung cancer” is false....”
Again, I refer you to the first citation, which also underscores the fact the line between “weak” and “strong” is done by fiat, another challenge to the so-called link between smoking and cancer.
The Japanese smoke more (if not the most) than most cultures yet are also one of the most healthy cultures. This goes to your “slightly more interesting question,” but it also goes the challenges of positively correlating smoking with “negative” outcomes. A further problem is that “negative outcomes” are normatively tied to cultural standards. Another problem is with average life expectancy comparisons, as they are to sensitive to outlier inflation.
Sorry, can you be more specific? Where does anybody claim that smoking is not strongly correlated with life expectancy?
The second “in house” link is a very simple thought experiment to explain the concept of confounding. It is meant as an example where evidential decision theory fails. In this situation, causal decision theory gives the right answer, it is certainly not a 50-50 proposition. Moreover, the correct answer within the thought experiment is that smoking does not cause cancer. This is because they postulated the existence of a deterministic confounder. This has no implications for whether or not such a confounder exists in the real world.
Because the confounders, ie the “smoking lesions”, would have to be unrealistically strong to fully explain the observed correlation between smoking and lung cancer. This is the part where I showed you the sensitivity analysis.
Of course we don’t have a “sure” way of knowing about causal relationships. But if you adopt “certainty” as your epistemic standard, you wouldn’t even be able to tell whether parachutes save lives in people who are falling from airplanes.
This is called an “ecologic” argument, and it is considered very weak. Note that your sample size is essentially 2, as the units you are making inferences about are countries, not individuals.
Now you’re just trolling… We’re talking about life expectancy, lung cancer, heart attacks etc here.