So That's the State of Epidemology?

Correlation not causality: the gift that keeps on giving

Eric Crampton links to a new study called "Lung cancer risk in never-smokers: a population-based case-control study of epidemiologic risk factors" by Darren R. Brenner and seven other authors (html full text with link to pdf for those who prefer the latter). Their main findings:
Any previous exposure to occupational exposures (OR total population 1.6, 95% CI 1.4-2.1, OR never smokers 2.1, 95% CI 1.3-3.3), a previous diagnosis of emphysema in the total population (OR 4.8, 95% CI 2.0-11.1) or a first degree family member with a previous cancer diagnosis before age 50 among never smokers (OR 1.8, 95% CI 1.0-3.2) were associated with increased lung cancer risk.
They also find a positive, but insignificant, association between exposure to second-hand smoke and lung cancer. More on those occupational exposures, though:
Occupational exposure to any of the putative lung carcinogens was associated with lung cancer risk in the total population (OR 1.6, 95% CI 1.4-2.1). Among never smokers, the odds ratio for exposure to any of the putative carcinogens was 2.1 (95% CI 1.3- 3.3). Specifically among never smokers, exposure to solvents, paints or thinners conferred an OR of 2.8 (95% CI 1.6-5.0), while exposure to welding equipment conferred an OR of 3.4 (95% CI 1.1-10.4) and exposure to smoke, soot or exhaust (other than tobacco) conferred an OR of 2.8 (95% CI 1.4-5.3). We did not observe significant associations for exposures to asbestos, pesticides, grain elevator dust, and wood dust among never smokers in our study.
This is a case-control study, which is a somewhat funny design. You take a large set of people and match up people who differ on your outcome of interest (such as lung cancer) but are the same, or at least similar on qualities you want to control for statistically because you are not interested in the possible effects of these qualities, but they may influence the outcome. You do not match on qualities the effects of which you want to measure (e.g., second-hand smoke). To be able to draw any inferences about the measured cause, such as second-hand smoke, to be the actual cause of any differences, you want to be really, really thorough in your matching procedure. It would be a shame if you falsely concluded that second-hand smoke casues lung cancer just because exposure to second-hand smoke is correlated with something else that you have not controlled for and which is the actual culprit.

Brenner et al. match on sex and ethnicity and also report adjusting for smoking, age and education (don't ask me how these adjustments work, but I assume they do). If I am reading the authors correctly, they analyse one risk factor at a time and hence do not, say, adjust for passive smoking when looking at, say, welding, but I'm no expert in case-control studies, so I'm probably only overlooking something.

Anyway, this is only one study. The real truth lies in the meta-analyses, and Brenner et al. cite one on passive smoking, the 1997 "The accumulated evidence on lung cancer and environmental tobacco smoke" by AK Hackshaw et al. in the prestigious BMJ. They look specifically at the effect of having lived with a smoker on never-smokers. Their main result:
The excess risk of lung cancer was 24% (95% confidence interval 13% to 36%) in non-smokers who lived with a smoker (P < 0.001). Adjustment for the effects of bias (positive and negative) and dietary confounding had little overall effect; the adjusted excess risk was 26% (7% to 47%).
As usual, though, you need to look at the details (p. 982; references omitted):
We used odds ratios unadjusted for potential confounding factors except in four studies, in which only adjusted estimates were available. For the cohort studies we used the published age adjusted relative risks (and 95% confidence interval).
That's right: They deliberately selected the shoddiest possible measures except when they absolutely couldn't help accepting measures adjusted for age. Unadjusted measures would be fine if researchers had matched on everything that may be relevant, but given that it appears to be customary to not even match on age, I have my doubts that every single study they analysed matched on passive welding et al. The adjustments for the effects of bias alluded to above are only for misclassification of ever-smokers as never-smokers and passive smoking due to sources other than the spouse.

Talking of bias, a reanalysis of the same sources by Copas and Shi that adjusted for publication bias reduced the risk estimate by 40%. Lord knows what the Hackshaw et al. findings mean, but my reaction certainly wouldn't have been, "Strong public health action is long overdue". Then again, I'm not editor of Tobacco Control.

Don't get me wrong. Passive smoking may well cause lung cancer, it's plausible enough. There may be newer meta-analyses using better methods. I'm not really all that interested in the topic, frankly. But this little exercise has exactly increased my confidence in what you can read in med journal abstracts and the quality of the products of the medical research community as a whole. It's regrettable; after all, the topics these people study are rather important.

Next: Global warming. Just joking.

1 comment:

pj said...

Obfuscatory and opaque controls for confounding factors pretty much characterises medical research, particularly epidemiological.

But since the structure of medical research incentives is pretty much designed to provide perverse incentives to manipulate results it isn't exactly surprising.